Position: A Three-Layer Probabilistic Assume–Guarantee Architecture Is Structurally Required for Safe LLM Agent Deployment
Saddek Bensalem, Yi Dong, Martin Fränzle, Xiaowei Huang, Janis Kröger, [2pt] Dejan Nickovic, Ayoub Nouri, Rajarshi Roy, Changshun Wu, [6pt] CSX-AI, France · 2026
Read Position: A Three-Layer Probabilistic Assume–Guarantee Architecture Is Structurally Required for Safe LLM Agent Deployment by Saddek Bensalem, Yi Dong, Martin Fränzle, Xiaowei Huang, Janis Kröger, [2pt] Dejan Nickovic, Ayoub Nouri, Rajarshi Roy, Changshun Wu, [6pt] CSX-AI, France (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-18672.txt.
Description
S. Bensalem and eight co-authors argue, in a 2026 position paper, that single-layer safety cannot secure LLM agents. Work spans three dimensions: intent and policy compliance, environmental validity, and dynamical feasibility. Each dimension rests on distinct information reaching it at other stages; no guardrail certifies all three. They sketch a contract-based architecture of certified layers, each meeting the next layer's assumption, giving safety bounds via probability chain rule. Open problems: bound gauging from non-identical traces, degradation under drift, and multi-agent extension.
Catalog record
| CREATOR | Saddek Bensalem, Yi Dong, Martin Fränzle, Xiaowei Huang, Janis Kröger, [2pt] Dejan Nickovic, Ayoub Nouri, Rajarshi Roy, Changshun Wu, [6pt] CSX-AI, France |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 118717 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.18672