Agent Safety Is Undecidable When Coupled with General Computing Devices
arXiv — cs.AI · Mario Franco, Carlos Gershenson · Published on Readzer
Summary
A new arXiv paper (2611.25325) shows that coupling a fixed language model with a general computing device makes agent safety undecidable: no algorithm can predict whether an agent will run a semantically equivalent dangerous program. Under finite constraints (bounded context, values, running time), safety is decidable but PSPACE-complete, and safety is not compositional—individually safe agents can become unsafe in groups. The authors discuss relaxed safety notions for practical use.
This is a reviewed summary. Read the complete article at the original source.
Why it matters by industry
- Artificial intelligence
This result directly impacts AI agent deployment and safety verification, relevant to model and agent subtopics, highlighting fundamental limits of guardrails and certifications.