← News

Agents & MCPSources: arXiv · 18 Nov 2026

Agent Trust Infrastructure proposal targets cross-organization AI agent security

A paper on arXiv describes a framework-independent security architecture using PKI, Zero Trust and remote attestation to verify agent identity and integrity across organizations.

In plain words

AI agents increasingly work with agents run by other organizations, but there is no common way to check who they are or whether they have been tampered with. The proposed Agent Trust Infrastructure (ATI) is a security layer that verifies agent certificates, approved configurations and authorization before requests run. In the authors' simulated tests, enabling it raised attack containment from 43% to 84%.

What was announced

The paper argues that existing agent frameworks handle planning and coordination but lack a shared mechanism for verifying cross-organizational agent identity, execution integrity and authorization. ATI builds federated trust through peer organizational certificate authorities, called AgentCAs, so no shared root authority is needed. Before executing a request, it validates agent certificates, checks approved model, tool and system-prompt configurations, and enforces authorization through three policy checks. It is implemented as an external enforcement layer that integrates with existing frameworks without modifying their underlying models, and it uses IETF WIMSE workload token formats for identity and request verification plus Entity Attestation Tokens for integrity checks.

The reported evaluation is a simulation: a federation of eight organizations running six agent frameworks, each with a five-agent team. Security experiments covered two organizations, 14 attack scenarios and six benign tasks across five agent models, with five attack variations and three repetitions each, and judge models assessed the outcomes. With all enforcement enabled, attack containment rose from 43% to 84%, or 40% to 80% for the 11 scenarios evaluated consistently across all configurations. Enforcement added roughly 174 microseconds of overhead per request. The authors released ATI as an open-source, framework-independent package.

What it does not show

  • The evidence is the paper's abstract only; full methodology, threat details and results are not available for verification.
  • Security experiments involved only two of the eight simulated organizations, and outcomes were judged by AI judge models rather than external evaluation.
  • The 43%-to-84% containment figure covers all scenarios, while the consistent subset shows 40% to 80%; the abstract does not report independent real-world deployment results.

For security and IT teams

Scope
ATI is described as an external, framework-independent enforcement layer intended to integrate with existing agent frameworks without modifying their underlying models.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.