Two agents filling real gaps, taking the collection from 50 to 52.
fix-verifier
poc-validator proves a bug is real before it reaches a report. Nothing covered the other end: someone says it's fixed, and that claim goes untested. That's how the same vulnerability gets reported twice.
Replays the original proof rather than a weaker substitute, tests the vulnerability class instead of the exact payload string, sweeps sibling endpoints and older API versions, and confirms the legitimate feature still works. Verdicts are CLOSED / PARTIAL / NOT FIXED / REGRESSED / BROKEN BY FIX.
Ships with a fake-fix pattern table, because most failed remediations fail the same handful of ways: client-side-only validation, blocklist instead of allowlist, one endpoint patched of several, WAF shim, suppressed error, auth added but authz still missing, fixed in main but never deployed.
supply-chain-auditor
Covers what the build pulls in rather than what the team wrote. SBOM inventory including transitive dependencies, dependency confusion weighted on resolution config (--extra-index-url, scope-to-registry mapping) because that is what actually decides exploitability, typosquatting, install-time hooks, lockfile and provenance integrity, maintainer risk.
Findings rank by blast radius rather than CVSS. A low-severity issue in a package that runs in CI with deploy credentials outranks a critical one in a sandboxed leaf.
It will not publish to a public registry and will not execute untrusted package code. Proving a namespace is unclaimed is the finding; claiming it is an action with real-world consequences.
Install
/plugin marketplace add 0xSteph/pentest-ai-agents
/plugin install pentest-ai-agents@pentest-ai-agents
Both agents carry the mandatory scope-guard block and refuse to act on undeclared targets.