“Patterns mislead. Deviations reveal.”
THE ANALYST'S CREED — EVERY REQUEST IS MATCHED AGAINST WHAT IS NORMAL.
“The AI said so.” is not a report.
Here is the replay. Here is the hash. Run it yourself.
$ bf verify 042 --replay
→ GET /api/v1/admin/users (session A → tenant B)
← 200 4,113 bytes [{"id":…}] ✓ matches evidence hash
✓ CONFIRMED · IDOR · HIGH · cvss 8.1 sha256:9f4c…a2e1A buying-committee unlock — not a feature.
Across live hunts on public programs. Every number above is a claim the engine will replay for you.
What is AI penetration testing?
AI penetration testing uses autonomous agents to attack a target the way a human security researcher would — reconnaissance, fingerprinting, exploitation, chaining — but continuously, in parallel, and at a fraction of the cost. BountyForge is an AI pentest engineer: it plans its own tests, executes them in an isolated sandbox, and writes the report.
Can an AI really find real vulnerabilities?
Yes — with a verification rail. BountyForge never ships a guess: every confirmed finding must survive nine layered verification gates, including replaying the proof-of-concept against the live target. A finding that does not re-execute is demoted, not reported.
How is BountyForge different from a vulnerability scanner?
A scanner runs fixed templates and returns noise. BountyForge senses each target’s terrain, spends its budget where evidence says findings live, learns what worked on similar targets across hunts, and chains individual weaknesses into exploit paths — the judgment layer scanners do not have.
Does it work for bug bounty hunting?
Bug bounty programs are a primary surface. The engine honors scope manifests (in-scope, out-of-scope, disallowed test types), rate limits, and authorization records on every request, and formats verified findings for HackerOne, Bugcrowd, Intigriti, and Immunefi submission.
Can it audit smart contracts?
Yes — Solidity, Solana, and Move targets are audited with static analysis, invariant checks, and fork-based PoC execution, so a reported exploit is proven against a pinned live block, not asserted.
Is autonomous pentesting safe to run on production systems?
It is authorization-gated by design: scans require a valid scope record declaring authorized targets before a single packet is sent, run inside isolated sandboxes, and stop at out-of-scope boundaries enforced at the tool level.