Study ties agent write safety to how lookups reveal competing records
Researchers apply linguistic presupposition theory to LLM agents, finding that whether a tool lists, counts, or hides a second record strongly changes how often agents act on false assumptions.
In plain words
When an AI agent cancels a booking or deletes a folder, it can act on a wrong assumption about what exists. The researchers tested how the design of the agent's lookup tools affects this. Tools that list every matching record, or that check a detail the user knows, led agents to pause and ask far more often than tools that only count results.
What was announced
The authors frame the problem with presupposition theory: an agent should act only when each claim behind an action is visibly supported by its tools, answered by the user, or disclosed while the action can still be undone. A simple bound shows that when a lookup cannot distinguish one record from two, any agent acts without asking as often with two records as with one. In tests where a search could list, count, or omit a second match, a count left most of 36 earlier agents acting on the record they saw, while listing both records stopped most of them. The newest Claude and GPT models always opened a counted second record and then never acted.
User statements changed the picture. When a user claimed a folder was empty, a count of its one item stopped agents about as often as a full listing, and surfacing the claim mostly cut mistaken actions when it was false. In a retail benchmark, adding a second customer with the same name and zip code meant a lookup returning that customer first led to wrong-account writes in roughly a quarter of conversations, while a lookup checking a street the user knew kept completion near the one-customer level. The authors conclude that safe writes need lookups that list or check competing records, and evaluations that vary what tools show.
What it does not show
- The evidence is the paper's abstract only; full methods, model versions, and benchmark details are not available to verify the reported figures.
For security and IT teams
- Scope
- Findings cover agent write actions such as cancellations, deletions, and account writes in the tested benchmarks, not general agent behavior.
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.