Attestation
A New Protocol Bets on Recomputation Instead of a Central Trust Authority
Most accountability proposals for cross-organization AI agents still assume someone has to vouch for the agent: a registry, a platform, a credential issuer. A recent paper, "Trust Without Trusting: A Recomputable Trust Protocol for Autonomous Agents," takes a different position. Rather than relying on a trusted third party to attest that an agent behaved correctly, it proposes that any party can recompute and verify an agent's trust state directly, without deferring to a central authority.
The paper's framing is useful because it maps where everyone else already stands. Singapore's IMDA governance framework, released this year, is described as requiring each agent to carry a verifiable digital identity and an audit trail of which agent acted under whose authorisation. NIST's new AI Agent Standards Initiative starts from the same gap: agents are commonly treated as generic service accounts without dedicated identity, authorization, or accountability controls. The EU AI Act's Article 14 obligations on human oversight apply to agents in high-risk domains, but the paper notes that duty falls on the relying party, not on the protocol itself — which is exactly the seam it's trying to patch.
On the industry side, Anthropic's recent papers on trustworthy and zero-trust agents argue that model-layer safety alone can't secure agentic systems and call for shared infrastructure no single vendor can supply. Google's SAIF 2.0 risk map flags rogue-action and over-permissioned-tool risks, pushing for well-defined controllers and observable actions. Microsoft's Entra Agent ID solves agent identity, but only inside a single enterprise's perimeter. Every one of these frameworks stops at the boundary of the organization that built it.
That's the gap this protocol is aimed at: holding the exercise of authority to account across a network of organizations that don't share a root of trust. Instead of a certificate authority or a shared registry verifying claims, the model trades on recomputation — any party with access to the right inputs can independently rederive the trust state and check it against what was claimed, rather than taking a signed attestation from an issuer at face value.
It's an early-stage academic proposal, not a shipped system, and the gap between a working paper and something agents actually run in production is wide. But it's a clean articulation of a problem that keeps resurfacing across this space: every accountability framework built so far assumes a trusted party sits somewhere in the loop, and none of them explain what happens when the agent, the tool, and the auditor all belong to different companies with no shared authority to appeal to.