A candidate is not a verdict
Scanner output and model suggestions identify things worth investigating. They do not, by themselves, establish that an exploit works. Pentest-AI keeps unsupported or unproven findings as candidates.
The oracle checks the claim
An oracle runs a specific verification recipe. It repeats the exploit and checks a control that should not demonstrate the same weakness. Only a passing result earns a verified verdict.
For a trusted-header bypass, that means privileged content must appear with the attack header and be denied without it. An application that returns HTTP 200 for everything is not sufficient evidence.
The model cannot award a verified verdict. Its role in the investigation is different from the oracle's role in confirming a finding.
Keep the evidence with the finding
A proof capsule records the finding, its verification recipe and the oracle's receipt. A supported capsule can be replayed against the recorded target.
Know the limits
- Verification describes the tested conditions, identity, route and application state. It does not establish that the whole application is secure.
- Unsupported bug classes can remain candidates even when they deserve manual investigation.
- Authentication expiry, target changes and network conditions affect replay.
- Capsules have an integrity check, but are not signed proof of authorship. Rerunning the check is the trust mechanism.