Back to blog

Why ML-DSA Matters for AI Agent Audit Trails

Mar 1, 2026

NIST published FIPS 204 in August 2024, finalizing ML-DSA as the post-quantum digital signature standard after a seven-year evaluation. For most software, the transition can be planned over the next decade. For AI agent audit trails, the timeline is more urgent.

What ML-DSA is

ML-DSA provides the same guarantees as RSA and ECDSA: authenticity, integrity, non-repudiation. The difference is the math. RSA and ECDSA rely on problems (integer factorization, discrete logarithms) that Shor's algorithm solves on a quantum computer. ML-DSA rests on the Module Learning With Errors problem, a lattice-based problem with no known efficient quantum algorithm.

Three parameter sets: ML-DSA-44 (Level 2, 128-bit), ML-DSA-65 (Level 3, 192-bit), ML-DSA-87 (Level 5, 256-bit). ML-DSA-65 is the parameter set most commercial deployments should land on.

Why RSA and ECDSA are not enough

The threat model is "harvest now, forge later." Adversaries collect signed audit records today and wait for quantum capability. Records signed with ECDSA today that must be valid in 2040 are on a collision course with quantum computing timelines.

The NSA CNSA 2.0 guidance is public acknowledgment of the same concern at the nation-state level: the timeline for moving signed records onto post-quantum algorithms is measured in years, not decades, when the records themselves have to stay verifiable for decades.

Key sizes and performance

Property ECDSA (P-256) ML-DSA-65
Public key 64 bytes 1,952 bytes
Signature 64 bytes 3,309 bytes
Verify speed Moderate Very fast
Post-quantum No Yes

Signatures are about 50 times larger than ECDSA, but for audit records stored in a database, 3.3 KB per signature is small. A million records add about 3.2 GB. Verification is faster than ECDSA on most hardware.

Why AI agents specifically

AI agent audit records have long retention requirements: EU AI Act requires system lifetime plus a reasonable period, DORA mandates at least five years, CNSA 2.0 requires full post-quantum transition by 2035, and financial services commonly retain records for seven to ten years.

ML-DSA addresses the problem at the source. Records signed today remain verifiable in 2036, 2046, and beyond. No migration to plan, no re-signing project to budget for.

How Asqav uses ML-DSA

Every agent gets a dedicated ML-DSA-65 key pair. Private keys are generated and stored server-side. The public key serves as the agent's verifiable identity. Signatures are standard FIPS 204 and can be independently verified by any compliant implementation: Asqav, liboqs, BoringSSL, or any future library.

For the full specification, see NIST CSRC. To see how Asqav implements ML-DSA-65, explore our documentation.

Stay ahead of AI compliance

Get practical insights on AI agent security and compliance obligations. No spam, unsubscribe anytime.