Skip to content

Security: nrcrews/inference-audit

Security

SECURITY.md

Security policy

Report vulnerabilities privately through GitHub's security-advisory form for this repository. Do not include live API keys, private prompts, or unredacted provider responses in a public issue.

Credential handling

The CLI loads .env from the current directory without overriding existing environment variables. API keys are read from the named variable and placed only in the outbound Authorization header. They are not written to JSONL, reports, or normal diagnostics. --header Authorization=... is rejected so a credential cannot enter a copied command or diagnostic path.

Output sensitivity

Observation files contain prompts and model responses. Those can include personal, proprietary, or regulated data when a custom endpoint or future custom suite is used. Protect, retain, and publish them according to the sensitivity of the inputs. The built-in suite uses synthetic prompts.

HTTP error bodies are truncated to 500 characters but can still carry provider-supplied detail.

Supported versions

Security fixes are released for the latest minor version. Until 1.0, upgrade to the newest release before reporting a behavior discrepancy.

There aren't any published security advisories