Policy

Security posture

Agent Blackbox is designed to create tamper-evident records for observable agent activity while keeping audit evidence separate from operational telemetry.

Evidence separation

Forensic audit evidence is treated as a distinct data plane from product analytics, health checks, and operational logs. Admin surfaces should rely on redacted previews and stable hashes instead of plaintext payload exposure.

Encryption by default

When evidence encryption is configured, observable audit payloads are encrypted before storage or remote custody. Remote and high-assurance modes should fail closed when required evidence-encryption material is unavailable.

Identity before evidence

AgentBox runs must resolve a registered organization, operator, and distinct agent instance before evidence starts. Missing identity should stop the run and provide registration guidance.

Verification limits

Agent Blackbox verifies observable behavior, hashes, signatures, timestamps, custody metadata, and exported evidence packages. It does not claim access to private model chain-of-thought or hidden cognition.