Pentest-AI

Pentest-AI connects an AI client or model to a penetration-testing workflow for applications you own or are authorized to test. The model helps drive the investigation, and a machine oracle decides whether a finding is verified. It runs as an MCP server for coding agents such as Codex and Claude Code, as a command-line tool with a local Ollama model or a cloud provider, as deterministic checks with no model at all, or as a gate in CI. Findings are stored locally in SQLite.

Verification works in three steps. A probe or scanner flags a candidate, a named oracle repeats the exploit and compares it against a control that should not show the weakness, and a verified finding records the oracle and its results in a portable proof capsule that can be replayed later. For example, a trusted-header bypass must return privileged content with the attack header and deny access without it. The model cannot award a verified verdict, and third-party scanner output stays unverified until an oracle proves it.

A verified verdict applies only to the tested conditions, and a scan with no verified findings is not a full security assessment. Proof capsules are unsigned, so rerunning the check is the trust mechanism. Model-provider calls, installed scanners, and out-of-band callbacks for blind vulnerabilities have their own network behavior, and callbacks can point to your own collaborator server or be disabled.