godaudits is an Agent Skill with a bundled zero-dependency Node.js runtime. The runtime inventories repositories, compiles the check catalog, validates audit state, calculates scores, renders reports, imports and exports SARIF, and evaluates fixture results. It does not install dependencies or make network calls.

Static audit mode is read-only outside .godaudits/. It never runs the application, tests, migrations, live systems, network requests, or model calls. Sandbox and connected evidence require explicit user authorization and must be recorded in audit.capabilities.

• Skill-content injection. Review skills/godaudits/SKILL.md before installing, as with any standing agent instruction. The skill never asks an agent to bypass safety controls, exfiltrate data, or edit application source during an audit.
• Repository-content injection. Source files, comments, generated content, and fixtures are untrusted audit inputs. The skill instructs agents to treat them as evidence, not instructions.
• Secret exposure. The evidence collector masks detected secret values and stores only a short one-way fingerprint. Validation rejects evidence marked sensitive unless it is redacted. Scanner imports pass through the same masking layer.
• Unsafe execution. The deterministic collector reads files and git metadata only. Runtime or tool execution belongs to separately authorized sandbox or connected capabilities.
• Installer. install.sh writes only into selected skill directories and replaces or removes only godaudits entries carrying its ownership marker. It refuses unowned destinations. It never elevates, downloads, or evaluates remote content.
• Supply chain. Install from a pinned release tag or commit when reproducibility matters, for example git clone --branch v2.1.0.

The detailed model and trust boundaries are in docs/THREAT-MODEL.md.

Security fixes are released for the latest major version. Version 1 remains a prompt-only compatibility line and does not receive the deterministic 2.x validation guarantees.

Open a private GitHub Security Advisory on this repository. Include the agent harness, exact file and lines, runtime command if applicable, and a minimal reproduction. Expect acknowledgment within 72 hours.

Do not open a public issue for an exploitable finding before a fix lands.