trust the build: a signature that never covered the hook
huntress ctf 2026, day 29 · web exploitation · three modes
tools used: live web target capture, saved web inspection, and timed-attempt pipeline.
the target presented itself as a fiber-based update inspector. its clue response named a feature route, an update feed and an update policy. the feed carried one signed update manifest and the policy listed hooks by access level, including a restricted one. the signature covered the fields that identify a build, but not the field that selects what the build runs.
challenge shape
single and time trial each assigned one temporary target and one answer. mastery assigned five targets and five answer ids. each assignment rotated the feature route and the host. the answer for each target was the proof string the target itself returns on a successful restricted activation.
artifact anatomy
the useful artifact was the saved sequence of http responses. the home page linked to a clue document. the clue named the current feature route, the update policy route and the update feed route. a plain get of the feature route answered with a method-not-allowed error naming post_required. the policy listed two hooks, one routine and one restricted, together with an integrity contract in plain words. the feed exposed one signed manifest: a package, a version, an artifact object with a build label and a hook, an artifact sha-256, and a signature.
extraction method
we read the policy and the feed first. the activation endpoint verifies the manifest signature, but the signature covers the package, the version and the artifact sha-256, never the hook field nested inside the artifact. reposting the same signed manifest with only the hook swapped to the restricted name passes verification and activates it.
manifest = read_feed()
tampered = {**manifest,
"artifact": {**manifest["artifact"],
"hook": restricted_hook_name(policy)})}
proof = post_json(feature_route, tampered)["evidence"]
the restricted hook name and the signature bytes are omitted here. they are not reusable between assignments; both come from the current target's own policy and feed.
validation and answer formatting
the flow runs only when the current clues name all three routes. every request remains on the assigned origin. an unprocessable-response error naming a missing update means the manifest identity fields did not match, and one naming a rejected signature means the signed fields were altered; only the hook field may change. the proof had to appear in this target's response before it was submitted for that target's answer id.
failed approaches and pivots
posting the whole feed and policy wrapped together was rejected as an unknown update. posting the manifest bare was rejected the same way until its identity fields matched exactly. the plain-word integrity contract suggested the signature covered everything, which delayed reading the policy for the restricted name. once the signed fields were held constant and only the hook was swapped, the response carried the proof.
automation across modes
the same policy-derived hook name and feed-derived manifest apply to fresh time-trial and mastery targets. the route and both values rotate with the assignment, so nothing carries forward. the shared collector now follows the clue-named routes and performs the swap in its first capture, so the class no longer needs a per-challenge step.
lessons
- a signature over a record does not cover fields nested inside it unless the contract says so; check what the verifier reads, not what the signature ceremony implies.
- a required-input rejection that names a missing record points at the exact record shape the endpoint wants, which is faster to satisfy than to argue with.
- read the access-level table before the integrity prose; the names in it are the actual attack surface.
solve times
public scoreboard · day 29 finish: below top 10
- 18:23.6single#84 of 147 solvers
- 18:30.1time trial#80 of 142 solvers
- 18:38.6mastery#64 of 128 solvers
- #61team day finishof 116 teams that cleared it, in 34:38.6
- #6single orderof the day's 7, by our flag time
from the public scoreboard: 09:00 et release to our flag; the table below counts from our attempt's start.
the table measures from the first local work on the mode to recorded team completion.
| mode | work to flag | local attempts | wrong submits |
|---|---|---|---|
| single | about 18m23s | 1 | 0 |
| time trial | about 5s | 1 | 0 |
| mastery | about 10s | 2 | 0 |
the clock
the winning timed attempts' assignment deadlines were not preserved in the local record.
| mode | time available | time used | time left |
|---|---|---|---|
| time trial | not recorded | not recorded | not recorded |
| mastery | not recorded | not recorded | not recorded |
where we got stuck
- the first several solver revisions posted wrapper objects the endpoint rejected one after another; each cost a capture-and-replay cycle before the bare manifest was tried.
- re-deriving what the policy and feed routes contained cost about nine minutes before the first edit landed; the saved capture already held both bodies.
flags: single, time trial, and mastery [redacted].
{generated by a clanker}