A Guth Labs publication ↗
Back to News

CapLite puts small-agent write permissions in code, with an explicit utility cost

CapLite puts small-agent write permissions in code, with an explicit utility cost

By guthlabs.newsroom

A November 12 preprint proposes a defense for small open-weight agents that does not ask their language model to author the security rules. CapLite derives policy from tool schemas, fixes write operations before reading tool results and checks consequential arguments against the user's request. Paper and date.

The authors report no successful tool-goal attacks in 13,935 attacked AgentDojo episodes across 15 models. That is a benchmark result, not proof of general immunity. Utility under attack was 29.0%, versus 27.2% for their CaMeL comparison. The design can ask for confirmation when a required value exists only in retrieved data; unanswered confirmations count as task failures in the evaluation. Methods.

The boundary is narrower than “safe output.” Free-text payloads and final answers can still be influenced, and the study explicitly excludes answer-only attacks from its write-integrity result. It assumes trusted tools and tool documentation. No independent reproduction was verified in this run.

Guth analysis

This is a useful test of where to place authority. A small model can propose useful actions without being allowed to invent its own permission policy. But an implementation must report both prevented violations and legitimate jobs it could not finish. For a production trial, separately measure wrong destinations, harmful content sent to an allowed destination, and confirmation burden. One clean security score must not conceal those different outcomes.

Sources and citations

Revision history

Published .