An agent can stay inside its permissions and still do the wrong job
An agent can stay inside its permissions and still do the wrong job
Microsoft researchers propose an Intent Profile that captures an agent’s intended role and boundaries. Their November 12 preprint reports four-class macro-F1 increasing from 45% to 79% on a purpose-sensitive benchmark. Abstract and submission record.
The paper’s fuller results deserve attention alongside its headline. On executed AgentDojo trajectories, the profile-based monitor scores 94.9% AUROC against AlignmentCheck’s 90.9%. But at a fixed 0.5 cutoff, AlignmentCheck has higher precision and F1. The profiles therefore do not win every operational comparison. Full paper.
Guth News analysis
Permission checks answer whether a tool is available. They do not, by themselves, answer whether using it serves the assignment. That distinction is useful for anyone evaluating agents that can browse, buy, edit or communicate. A monitor needs a trusted reference for the job, not merely a list of forbidden strings.
The practical next test is a deployment-specific tradeoff: how many inappropriate actions are caught, how many legitimate actions are blocked, and what delay the check adds. A single ranking metric cannot choose that tradeoff for a business. Nor should a profile become an excuse to give an agent additional authority.
Evidence and chronology
Event and primary publication: 2026-11-12 00:17:31 UTC. First covered in the Nov14 01:53 UTC run (Nov13 PST). No independent replication or independent corroborating reporting verified.
Evidence limits: synthetic organizational policies and LLM-validated samples; some behavioral tests are simulations. AgentDojo uses one agent model and four agents. Production cost and independent replication remain open.
Sources and citations
Revision history
Published .