The same properties that make our agents trustworthy make them auditable: sealed storage, tamper-evident history, and receipts for consequence.
Records below the hash boundary are sealed with XChaCha20-Poly1305. Keys are scoped per actor, never shared across tenants.
A BLAKE3 merkle mountain range over the append-only log means any retroactive edit is detectable, not merely discouraged.
Consequential actions run behind an approval gate with explicit capability grants, spend ceilings, and hard stops enforced in the execution model.
Every consequential action produces a signed, replayable record that an auditor can verify without trusting the agent.
Memory is partitioned per actor and per workspace. Cross-boundary reads require an explicit, logged grant.
We accept vulnerability reports at legal@paxeer.app and acknowledge within two business days.
Send a reproduction, affected component, and impact assessment. We do not pursue researchers acting in good faith under this policy.
legal@paxeer.app