ATI lifts attack containment for agents across companies to 84%
With all checks on, attack containment went from 43% to 84%, and each request needed about 174 microseconds more.
Researchers made Agent Trust Infrastructure (ATI), a protection layer for AI agents that work across different organizations and use the tools of each other. Before a request goes on, ATI checks the agent certificate, the approved model, tool and system prompt, and 3 policy checks. It is external to the agent frameworks and changes no model. The researchers released it as open source.
How it works
Each organization has a certificate authority, a group that gives each agent a digital identity. No shared root authority is necessary. The layer checks these items before it lets a request go on:
- the certificate of the agent
- the approved model, tool and system prompt
- 3 policy checks for authorization
The numbers
The simulation had 8 organizations, 6 agent frameworks and 14 attack scenarios. Containment went from 43% to 84%. For the 11 scenarios that all configurations had, it went from 40% to 80%.
The limits
The tests were in a simulation, and a panel of judge models gave scores to the results. An 84% rate means that some attacks go through.
Insight: Many groups are making the identity check for agents
On 14 October 2026, 4 companies showed an open Chain of Trust that uses internet standards. Socure has a trust score and a connection from an agent to its owner. ATI has no shared root authority. This reads as a move to control the gate between agents, since the first check that companies accept can set the rules. Monitor which one large platforms accept.
Sources
Posted