We audited 25 major AI agents for one thing: do they produce cryptographic proof of what they actually did?
Not logs. Not conversation history. Not vendor-controlled records. Cryptographic receipts. Signed at execution time. Independently verifiable. Tamper-evident.
The answer, across every agent we tested, is no.
The detection problem is a receipts problem. An agent that produces no signed record of its actions takes 14 hours to detect because you're hunting behavioral anomalies. An agent that produces a bilateral receipt for every action can be verified in real time.
None of the 25 agents we tested close this gap.
Each agent was evaluated against five criteria. Scores reflect what exists in the agent's default behavior, not what could theoretically be added.
EU AI Act Article 12 enforcement begins August 2, 2026. These agents will need to produce records that meet evidentiary standards. Currently, none do.
Nobulex is the open receipt layer. Ed25519 signed. Bilaterally attested. MIT licensed. Already merged into Microsoft's Agent Governance Toolkit.
View on GitHub