CapLite blocks prompt-injection writes without model-written security artifacts
Original title: CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts
No attack succeeded in 13,935 attacked AgentDojo episodes, with utility under attack of 29.0% versus CaMeL's 27.2%.
The paper finds small language models cannot reliably write the security artifacts that enforcement defenses such as CaMeL, Progent, IPIGuard and DRIFT depend on. Across 15 open-weight models, mostly between 1.7B and 12B parameters, model-written policies, graphs and constraints admitted attacks, and CaMeL programs failed to run in up to 68% of episodes. CapLite derives policy from tool schemas and allows a write only when every argument that decides its effect is stated in the user's request.
- Model-written artifacts admitted attacks on 14 Progent, 13 IPIGuard and 11 DRIFT models.
- Thirteen of the 15 models tested have between 1.7B and 12B parameters.
- CapLite used 17.1k tokens per episode, 18% of CaMeL's, the highest utility per token.
- Injected text cannot add a write or change recipient, amount, object or destination.
From the original article
1 min read at arxiv.orgEnforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be trusted to write these artifacts. Across 15 open-weight models, 13 of them between 1.
By Zimo Xu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.