Please report security issues privately rather than opening a public issue. Use GitHub's Report a vulnerability (Security tab -> Advisories) so the report stays confidential until a fix is available. Include steps to reproduce and the affected version. We aim to acknowledge reports promptly and coordinate disclosure.

Sentinel is offensive-security tooling — only use it against systems you own or are explicitly authorized to test. The full ethics and ground rules are in RESPONSIBLE_USE.md.

Reports about the tool itself are in scope: for example, the Nexus agent bypassing its own guardrails, the audit chain failing to detect tampering, secret redaction missing a credential class, or a command-injection path in the CLI.

• Guardrails — the agent enforces .nexus/policy.json (protected paths, denied commands, per-turn write limits, secret-write blocking) before any file or command action, plus an org-level floor that a local config can only tighten.
• Destructive-command preflight — shell commands are classified and blocked or flagged (classifyDanger).
• Secret handling — files are scanned for credentials before writing, and text is redacted before it is sent to a cloud engine.
• Tamper-evident audit — enforced actions are appended to a hash-chained .nexus/audit.jsonl; run sentinel audit verify to check its integrity.

Verify your setup at any time with sentinel doctor.