← News

Agents & MCPSources: arXiv · 16 Nov 2026

Soteria pairs verified planning with runtime enforcement to keep LLM agents safe and useful

The arXiv paper reports that a formally verified plan, checked before any tool call, guides agents when unsafe actions are blocked, removing the usual trade-off between safety and task success.

In plain words

AI agents that use tools can break rules while trying to finish a task, and simply blocking rule-breaking actions can leave them stuck. Soteria has the agent write a plan that is formally checked against the rules before any tool is used, and that same plan then steers the agent when an unsafe action is blocked. In the paper's benchmarks, this raised task success while producing no policy violations.

What was announced

Large language model agents that chain multiple tool calls face two goals at once: finishing the user's task and staying within domain policies and correctness constraints. The authors of Soteria argue these goals are in tension, because blocking an unsafe action prevents a violation but can strand the agent with no principled way to recover, and they say existing approaches sacrifice one objective for the other.

Their framework combines verified hierarchical planning with runtime enforcement of specifications. Before execution, the agent produces a structured plan that is formally verified against the specifications prior to any tool invocation. During execution, the verified plan both keeps the agent on specification-conformant trajectories and provides guidance when unsafe actions are blocked. On benchmarks spanning a diverse range of tool-use tasks, the paper reports Soteria achieved 1.7 times the average task success rate of the best baselines with no violations of policy specifications. These are the authors' own benchmark results and have not been independently confirmed; the reported scope is the paper's abstract, so full experimental details are not available here.

What it does not show

  • Evidence is limited to the paper's abstract; the full experimental setup, benchmarks and baselines cannot be verified from the supplied passage.
  • The 1.7 times success-rate figure and the absence of violations are the authors' self-reported benchmark results, not independently confirmed.

For security and IT teams

Scope
The supplied source is the arXiv abstract only, so deployment details, threat model and evaluation methodology are outside the available evidence.

Sources

Source-support review completed 17 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 17 Nov 2026

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