Correction, 11 October 2026: this post first gave the EU AI Act high-risk dates as they stood before the amendment. Regulation (EU) 2026/1744, in force since 27 July 2026, moved the high-risk dates to 2 December 2027 for Annex III systems (originally 2 August 2026) and 2 August 2028 for Annex I systems (originally 2 August 2027). The text below uses the amended dates.
AI agents are in production, processing transactions, accessing customer data, making decisions. The security implications differ fundamentally from traditional software. If you cannot answer these five questions today, you have gaps that need attention before enforcement deadlines arrive.
1. "Who authorized this specific agent action?"
Traditional applications trace requests to service accounts and deployed code. AI agents break this model. They decide what to do at runtime based on context. Two identical agents given the same prompt may take different actions.
What good looks like: Every agent action is cryptographically bound to a verified identity via digital signature. When an auditor asks "who did this?", you present a signature that mathematically proves which agent acted.
2. "Can the audit trail be tampered with?"
Centralized log aggregators protect against accidental data loss. They do not protect against deliberate modification by insiders, compromised agents, or sophisticated attackers.
What good looks like: Every audit record is digitally signed at creation. Any modification to any field invalidates the signature. Tamper evidence is a property of the record itself, enforced by mathematics.
3. "Will our signatures still verify at the end of the retention window?"
AI agent audit trails must remain valid for years or decades under EU AI Act and DORA retention requirements. RSA and ECDSA signatures carry a built-in expiration date over windows that long, since advances in cryptanalysis (including quantum computing via Shor's algorithm) make them forgeable.
What good looks like: Audit records signed with ML-DSA (NIST FIPS 204), a NIST-standardised lattice signature. Records signed today remain verifiable across the long retention windows compliance regimes demand.
4. "Are we meeting upcoming compliance deadlines?"
EU AI Act, high-risk obligations from 2 December 2027 (Annex III; Digital Omnibus enacted as Regulation (EU) 2026/1744, in force 27 July 2026). Article 12 requires automatic logging. Article 14 requires human oversight. Article 9 requires risk management. Handling personal data or working in a regulated sector does not make an agent high-risk by itself. It is high-risk when it is used for a purpose listed in Annex III, such as recruitment or credit scoring, or when it is a safety component of a product covered by Annex I that has to pass a third-party conformity assessment.
DORA, already in force. Requires comprehensive ICT risk management, incident reporting, and resilience testing for financial entities.
What good looks like: Your infrastructure already generates the audit artifacts these regulations require: identity binding, tamper evidence, policy enforcement, and human oversight capability through verifiable technical controls.
5. "Can any single entity unilaterally authorize agent actions?"
If your agent has unrestricted tool access with no policy evaluation before execution, you have a single-party authorization model where the agent is the sole decision-maker.
What good looks like: Agent actions are subject to policy evaluation before execution, with multi-party approval for high-risk operations. The policy engine is separate from the agent. Approvals and denials are recorded as part of the signed audit trail.
Where to start
These are the questions auditors, regulators, and board members will ask. Asqav provides the infrastructure to answer all five: cryptographic identity, tamper-evident trails, ML-DSA-65 signatures, compliance-ready records, and pre-execution policy enforcement. Explore our documentation or get in touch.