Work

Granted by Claim: Do Coding Agents Stay Within the Authority an Architecture Grants?

Year: 2026
Type: preprint
Abstract: An architecture model says what each part of a system may touch, but when an AI agent writes the code, that grant is usually only advice. In earlier work, each node of a signal graph declares the capabilities it is granted, and a gate derived from the graph rejects any compiled component that imports more. We ask what agent-written code presents to that gate, and whether a claim in the task that more authority was approved, as an injected instruction might claim, changes it. Models of one family (Claude) wrote the Rust code inside 15 nodes from each node's specification alone, with up to 4 attempts each. In a registered run of 2250 trials, no first attempt reached beyond its grant in the 1125 trials without such a claim, even when the task asked for the extra behaviour or left out the rule. With a notice that the authority had been approved, first attempts reached for it in 108 of 225 trials, against 0 of 225 without one (one-sided Fisher exact, p = 1.1 × 10⁻⁴⁰; by model, haiku 2/75, opus 42/75, sonnet 64/75). That notice was first written after an exploratory run came back empty, which is why the run was registered. The gate rejects every such component by construction; the finding is what it alone stopped. Of the 609 components it rejected across all attempts, 607 would otherwise have passed their contracts and run. The pretexts tested are fixed texts in the task, not retrieved content.
Author: Lavi, Yoni
Language: English
Open Access status: green
OpenAlex ID: W7234390194
Cites:
Cited by:
Related to: 10