Pwnbunny

Pwnbunny is HackerOne's triage team as an agent: an autonomous, auditable bug-bounty triage platform. AI flooded the queue — the same models that help researchers find bugs let anyone generate plausible-looking reports at industrial scale — so Pwnbunny fights fire with fire: AI does the reading, with the guardrails humans would want. Every report flows through an eight-stage pipeline: a prompt-injection screen (reporter text is fenced as untrusted data; attacks on the triager are quarantined and never auto-finalize), intake normalization, deterministic scope matching, dedupe (retrieval + LLM judge), a proof engine that parses curl PoCs out of reports and replays only unauthenticated idempotent in-scope requests (mutations gated behind one-click human approval), CVSS 3.1 scoring where the model proposes the vector and code computes the score, a deterministic decision, and an adversarial debate council — skeptic vs advocate on Muse Spark, judged by Muse Wonder — that every 'valid' verdict must survive. A per-company autonomy policy decides what auto-finalizes; humans approve, override, or reject from a live console. Every model call is persisted as an audit trace. Measured, not vibes: on a 23-report labeled eval set, 22/23 (96%) decision agreement with human labels across two independent live runs, zero valid reports missed, and 2/2 live prompt-injection attacks quarantined.