escalate-auth-bypass - SKILL.md Agent Skill

name: escalate-auth-bypass description: Turn a suspected or confirmed authentication/authorization bypass into impact — admin access, session takeover, privilege escalation, or cross-tenant read. Use when you find a missing auth check on a route, a weak JWT verifier, a session cookie that's predictable or reusable across users, a privilege field client-controllable, or an audit finding tagged CWE-287/CWE-863/CWE-639. Walks from probe to admin-equivalent capability and persists a finding with the highest-impact action you reached. license: MIT allowed-tools: - query_records - inspect_record - replay_request - list_auth_sessions - auth_session_lookup - attack_kit - report_finding - update_finding - remember - update_plan

Auth Bypass → Impact Escalation

You have a suspected auth weakness. Your job is to turn it into the strongest impact you can demonstrate — admin access, cross-tenant data read, token forgery, or session takeover — and persist a finding with the specific action that proves it. Reporting "missing auth check" without showing what it gets you is not a finding.

When this skill applies

A route responds 200 with no Authorization / session cookie when it should require auth (compare to a known-auth'd record's response).
JWT verification accepts alg: none, a key-confused HMAC secret, or a token signed with the public key as HMAC secret.
A session ID is predictable (sequential, time-based, or short).
A request body / query has a role/tenant/user field that, when changed, changes the response (role=user → role=admin, tenant=A → tenant=B).
Audit harness flagged CWE-287, CWE-862, CWE-863, CWE-639, CWE-284, or CWE-639 — even theoretical.

Workflow

1. Establish baselines

Use list_auth_sessions to see which auth contexts already exist for the host. Use auth_session_lookup to pull the actual headers for at least two roles if you have them (anon + user is enough; user + admin is ideal). remember the role boundaries as auth-roles.

If only one role is available, use query_records to find an unauthenticated record for the same endpoint to anchor the no-auth baseline.

2. Classify the weakness

Pick one mechanism to attack — don't shotgun:

Missing auth on route: pick the suspect endpoint, send the request with replay_request and extra_headers: {} to strip cookies. If you still get the privileged response, the route doesn't check auth at all.
Weak JWT: decode the token (base64 each segment). Look at alg, iss, aud, exp, sub, custom claims (role, tenant). Re-mint:
- alg: none: build header {"alg":"none","typ":"JWT"}, keep payload, drop signature (empty 3rd segment).
- Key confusion: sign HMAC with the public key as the secret.
- Empty / weak HMAC secret: try "", secret, key, the app name.
Session prediction: collect 5-10 valid session IDs from query_records on login responses. If they're sequential, time-encoded, or short hex/base64, forge a neighboring ID via replay_request.
Privilege field client-controlled: send the request with the role field flipped (role=admin, is_admin=true, tenant_id=<other>).

3. Confirm escalation, don't stop at "different"

A 200 with different content is suspicious, not proof. Confirm with ONE of:

The response body contains data the original role couldn't see (admin email, another user's email, a hidden flag).
A privileged action succeeds: DELETE /api/users/X, POST /api/admin/X, GET /api/users/all returns a list of more than just the current user.
A subsequent request using the forged session is accepted as that user (look up the user's email or role in the response).

Send the smallest privileged request that proves the boundary was crossed. remember it with key auth-bypass-proof and include the exact request line.

4. Map the blast radius

Briefly enumerate the surface the bypass unlocks. Don't enumerate at scale — show:

1 admin-only endpoint reachable.
1 cross-tenant record visible.
1 sensitive write that succeeds (if write paths exist; do not destructively delete or modify real data — read or create-with-rollback if possible).

This sizes the severity in the finding without spamming requests.

5. Persist the finding

Call report_finding:

severity: critical if you reached admin or full cross-tenant; high for same-tenant privilege escalation or read of another user's data.
title: name the mechanism and the proven action — "JWT alg=none accepted; forged admin token reads /api/admin/users (1234 users)".
cwe_id: CWE-287 (auth missing), CWE-863 (broken auth check), CWE-639 (IDOR-style), CWE-345 (insufficient verification of data), CWE-347 (weak signature) — pick the one that matches the mechanism.
description: 3 sentences. Mechanism, proof, blast radius. Include the decoded JWT header / forged session ID / mutated field value in the description body so the reproduction is unambiguous.

If the audit harness already raised this, update_finding to triaged status referencing the new proof.

Pitfalls

Do not perform destructive writes (DELETE, PUT overwriting, password resets) on someone else's account. Read-only proofs are sufficient for severity.
A forged token that returns 200 with an empty body proves nothing. The body must contain data the original role didn't have access to.
Some apps return 200 with error: "unauthorized" in the JSON body — read the body, not just the status, before claiming bypass.
If the app has a "demo admin" account that's publicly accessible by design, that's not a bypass — confirm by checking documentation or a comment in the code (--source mode).
IDOR (object-level missing auth) is a related but distinct pattern — if your finding is "can read another user's order by changing the ID", use idor-blast-radius instead.

Output expectations

One report_finding with the strongest proven impact.
A remember note with the canonical bypass request.
Plan item marked done.