Case study: pre-2FA token timing harness
Engagement: authorized 2FA-bypass research on a large commerce platform. Role: independent security researcher. Deliverable: a stdlib-only Python tool and a repeatable test procedure.
The problem
The interesting question in 2FA bypass work is not "can I guess the code." It is: is the session or token issued before the second factor already accepted by authenticated endpoints? If yes, the second factor is decorative for anyone who can capture that first-factor token.
Answering it safely needs discipline: one request per endpoint, no retries, stop on rate-limit, and never touch another user's data.
Approach
Built a small tool with four subcommands:
| Command | Purpose |
|---|---|
cookies | inspect a Netscape cookie jar for the session shape |
probe | one GET per endpoint, 2 s apart, stop on 429; report accepted/blocked |
token | offline decode (length, prefix, JWT claims if present; no signature check) |
compare | decode both a pre-2FA and post-2FA token, diff scopes, and probe identity with each |
The verdict matrix for compare:
PRE_2FA_TOKEN_ACCEPTED— the first-factor token works → critical.2FA_ENFORCED_AT_ISSUANCE— pre-token rejected, post-token works → correct.BOTH_REJECTED— neither works → no conclusion.
What was delivered
pre2fa_probe.py— dependency-free, rate-limited, prints the served identity so a token for the wrong account is caught instead of silently proving nothing.- A written procedure: capture the token at the moment the app would prompt for the second factor, then compare it to one issued after.
- Fixture-verified logic for both the accepted and gated paths.
Result
The supplied test account no longer resolved at the identifier step, so the live run is blocked on access, not on technique. The tooling is validated and runs the moment an authorized account exists.
Why this matters to a client
- You get a safe, repeatable harness instead of ad-hoc probing that risks lockouts or rate-limit bans.
- The pre/post comparison answers the real question the program asks, with evidence rather than a hunch.
- It runs offline for token shape and needs only a handful of requests live.
Testing is limited to accounts the client authorizes; the tool stops on 429 and never retries.