Franco, Mario, Gershenson, Carlos · Nov 19, 2026
Franco and Gershenson: Why Agent Safety Cannot Be Verified
Markdown · JSON · Original source
Mario Franco and Carlos Gershenson argue that coupling a fixed language model with a general computing device makes agent safety undecidable. They say no algorithm can predict whether an agent will run a program semantically equivalent to a given dangerous program. On that basis, they describe guardrails, blacklists, whitelists and certifications as imperfect safety solutions, not guarantees. The authors add that safety is decidable under finite constraints—bounded context, values and running —but remains intractable: the problem is PSPACE-complete even when the agent’s description is known. They also argue that safety is not compositional: agents that can be proven safe on their own can still produce unsafe behavior when grouped together. This means that, in their account, establishing safety for each agent separately does not establish safety for the group. The abstract says the paper discusses relaxed notions of safety that could be used in practice, but does not identify them or give their conditions.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 19, 2026 (2026-11-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-20T08:03:26.695Z. First observed: 2026-11-20T05:10:19.467Z. Retrieved: 2026-11-20T08:02:32.926Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract, not the full paper; it does not detail the arguments, assumptions, or proofs behind the claims. The abstract does not identify the relaxed notions of safety discussed or explain the conditions for using them in practice.
Open question
Which relaxed notions of safety do the authors discuss, and under what conditions do they propose they could be used in practice?
Source-stated
Franco and Gershenson argue that coupling a fixed language model with a general computing device makes agent safety undecidable.
We show that this coupling makes safety undecidable.
Source-stated
The authors say no algorithm can predict whether an agent will run a program semantically equivalent to a given dangerous program.
no algorithm can predict whether an agent will run such a program that is semantically equivalent.
Source-stated
The authors characterize guardrails, blacklists, whitelists and certifications as imperfect safety solutions.
guardrails, blacklists, whitelists and certifications can only lead to imperfect safety solutions.
Source-stated
They say safety is decidable but intractable under finite constraints, and the problem is PSPACE-complete even when the agent description is known.
Under finite constraints (bounded context, values and running time) safety is decidable but still intractable. Specifically, the problem is $\mathsf{PSPACE}$-complete even when the description of the agent is known.
Source-stated
The authors argue that individually safe agents can produce unsafe behavior when grouped together.
agents that can be proven safe on their own can still produce unsafe behavior when grouped together.
Source-stated
The abstract says the paper discusses relaxed notions of safety for practical use, without specifying them in the supplied text.
We also discuss relaxed notions of safety that could be used in practice.
Source and contributor context
Franco, Mario, Gershenson, Carlos · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d28e2f22-9c13-4ab6-8db6-adf6f25c4b60.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T08:03:26.695Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: efd67994a89d8c08f8e682f00512a26683497153ce359705fd4ba160fbe37adf. Source text hash: edf0c4185590e8f89174ab671e373d3ae30c2f5f9f298b2047f372cdeb04f6ea.
Revision and source-check history
Version 1 · 2026-11-20T08:03:26.695Z
Franco and Gershenson: Why Agent Safety Cannot Be Verified
Mario Franco and Carlos Gershenson argue that coupling a fixed language model with a general computing device makes agent safety undecidable. They say no algorithm can predict whether an agent will run a program semantically equivalent to a given dangerous program. On that basis, they describe guardrails, blacklists, whitelists and certifications as imperfect safety solutions, not guarantees. The authors add that safety is decidable under finite constraints—bounded context, values and running —but remains intractable: the problem is PSPACE-complete even when the agent’s description is known. They also argue that safety is not compositional: agents that can be proven safe on their own can still produce unsafe behavior when grouped together. This means that, in their account, establishing safety for each agent separately does not establish safety for the group. The abstract says the paper discusses relaxed notions of safety that could be used in practice, but does not identify them or give their conditions.
Original source
cs.AI · Article
Source details
Original source date: Nov 19, 2026 (2026-11-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-20T08:03:26.695Z. First observed: 2026-11-20T05:10:19.467Z. Retrieved: 2026-11-20T08:02:32.926Z.
Limits
The supplied evidence is the abstract, not the full paper; it does not detail the arguments, assumptions, or proofs behind the claims. The abstract does not identify the relaxed notions of safety discussed or explain the conditions for using them in practice.
Open question
Which relaxed notions of safety do the authors discuss, and under what conditions do they propose they could be used in practice?
Source-stated
Franco and Gershenson argue that coupling a fixed language model with a general computing device makes agent safety undecidable.
We show that this coupling makes safety undecidable.
Source-stated
The authors say no algorithm can predict whether an agent will run a program semantically equivalent to a given dangerous program.
no algorithm can predict whether an agent will run such a program that is semantically equivalent.
Source-stated
The authors characterize guardrails, blacklists, whitelists and certifications as imperfect safety solutions.
guardrails, blacklists, whitelists and certifications can only lead to imperfect safety solutions.
Source-stated
They say safety is decidable but intractable under finite constraints, and the problem is PSPACE-complete even when the agent description is known.
Under finite constraints (bounded context, values and running time) safety is decidable but still intractable. Specifically, the problem is $\mathsf{PSPACE}$-complete even when the description of the agent is known.
Source-stated
The authors argue that individually safe agents can produce unsafe behavior when grouped together.
agents that can be proven safe on their own can still produce unsafe behavior when grouped together.
Source-stated
The abstract says the paper discusses relaxed notions of safety for practical use, without specifying them in the supplied text.
We also discuss relaxed notions of safety that could be used in practice.
Source and contributor context
Franco, Mario, Gershenson, Carlos · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d28e2f22-9c13-4ab6-8db6-adf6f25c4b60.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T08:03:26.695Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: efd67994a89d8c08f8e682f00512a26683497153ce359705fd4ba160fbe37adf. Source text hash: edf0c4185590e8f89174ab671e373d3ae30c2f5f9f298b2047f372cdeb04f6ea.
passed · 2026-11-20T08:03:26.695Z · anthropic/claude-opus-5.5
Publication: d28e2f22-9c13-4ab6-8db6-adf6f25c4b60
attribution — pass: The reading names Mario Franco and Carlos Gershenson as authors and consistently frames each finding as their argument ('argue', 'say', 'in their account'). Nothing is presented as established fact or as Escape Meridian's view.
Authors:Mario Franco, Carlos Gershenson
summary — pass: The reading faithfully covers every abstract claim and keeps the authors' force: undecidability, the limits of guardrails, decidable-but-intractable under bounds, non-compositionality, and relaxed safety. 'Undecidable' is effectively explained by the next sentence, which says no algorithm can predict the behavior. Compositionality is explained in plain language. 'PSPACE-complete' and 'semantically equivalent' should get term-library tooltips if missing, but they do not block understanding. The question targets a real gap.
Finally, safety is not a compositional property, i.e., agents that can be proven safe on their own can still produce unsafe behavior when grouped together.
date — pass: The originalPublished date of 2026-11-19 matches the arXiv v1 submission date. It is not the later feed date of 2026-11-20.
[Submitted on 19 Nov 2026]
support — pass: Every claim is backed by a verbatim abstract excerpt. Calling these measures 'not guarantees' restates 'can only lead to imperfect safety solutions' without adding anything. The statement that separate per-agent safety does not establish group safety is labeled as the authors' account and follows from the abstract. No benefits, costs or forecasts are invented. The limitations correctly note that only the abstract was available.
In general, guardrails, blacklists, whitelists and certifications can only lead to imperfect safety solutions.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, and it is the source URL itself. It is the most direct way for readers to reach the original.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv