Skip to content

Security: mrwersa/agentverity

Security

SECURITY.md

Security and data handling

Reporting vulnerabilities

Please report a suspected vulnerability privately through GitHub's security advisory form. Do not open a public issue for an unpatched vulnerability.

Probe and output data

AgentVerity machine reports, progress events, and snapshots identify probe inputs by SHA-256 fingerprint rather than storing the raw text. Fingerprints are identifiers, not anonymisation. An attacker can test guesses against a fingerprint when probe text comes from a small or predictable set.

Snapshots retain the approved observation value because it is the reference being checked. A snapshot over the text layer may therefore contain model output with sensitive content. Recorded exception messages may also echo provider request data. Store these artefacts under the same controls as test fixtures and model logs.

The optional OpenTelemetry bridge emits aggregate agentverity.* attributes. It excludes raw prompts, outputs, fingerprints, majority-verdict values, relation names, and exception messages. Exporters, collectors, and monitoring backends remain outside AgentVerity's control. Review their retention and access policy before enabling the bridge.

Decision-suite JSON files are different. They are source test datasets and contain each raw input plus its intended decision. Treat them like test fixtures, not privacy-minimised reports. The generated JSON, JUnit, snapshot, and OpenTelemetry outputs do not copy those raw inputs.

Imported evidence also retains raw inputs and observations. Its provenance mapping is caller supplied and is not filtered. JSON, JUnit, and terminal diagnostics can echo observation values on finding paths, while JSON and terminal output retain recorded exception messages. The executable security and data-retention audit maps each surface and its caveats; “excluded” never means a provider, adapter, exporter, or caller-controlled field cannot log the same value elsewhere.

Concurrent execution

Concurrency is disabled by default. max_workers overlaps distinct inputs but never repeats for one input. Only enable it when the agent and its dependencies are safe for concurrent calls.

There aren't any published security advisories