Soteria pairs verified plans with runtime checks to keep AI agents on task
A November 13 research paper proposes a different approach to agent guardrails: give the system a checked plan it can return to when an action is blocked, rather than merely telling it to stop.
Soteria, by Mike He, Ankush Desai, Sharad Malik and Aarti Gupta, combines formal plan verification with enforcement during tool use. The authors report an average task-success rate 1.7 times that of their strongest baselines, with no violations of the policy specifications tested. These are the researchers’ results, not independent validation.
The full paper describes 207 tasks across three benchmarks and four models. Its setup includes human inspection of machine-generated specifications. If complete plan verification fails within the retry budget, the system can fall back to a partially verified plan while retaining runtime checks.
Analysis: why it matters. An agent that refuses unsafe actions but cannot finish legitimate work is a limited product. Recovery guidance is therefore worth evaluating alongside prevention. But formal conformance is only as useful as the rules and environment it models. This research should not be read as a guarantee that an arbitrary deployed agent is safe.
Evidence and timing. The arXiv submission is dated November 13, 2026, at 08:14:04 UTC. This is newly covered research, not a new release today. Independent replication was not located.
Sources and citations
Revision history
Published .