CapLite: Prompt Injection Defense Without Model-Written Security Artifacts

1. 01:07
   arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

   CapLite: Prompt injection defense for small language model agents that requires no model-written security artifacts

   The paper proposes CapLite, a prompt injection defense method for small language model agents that requires no security artifacts written by the agent model itself. Enforcement-based defenses such as CaMeL, Progent, IPIGuard, and DRIFT require the agent's own model to write programs, permission policies, dependency graphs, or constraint lists, which are then executed by code; the authors tested on 15 open-weight models (13 of which have 1.7B to 12B parameters) and found that the policies, graphs, and constraints written by the models made 14, 13, and 11 models respectively exploitable by attacks under Progent, IPIGuard, and DRIFT, and that CaMeL's programs failed to run in up to 68% of turns. CapLite derives policies from tool schemas, commits write calls before any tool result is read, and allows a write only when every parameter that determines the object affected by the write is explicitly stated in the user request, thereby guaranteeing write integrity: injected text cannot add write operations, and write operations cannot reach recipients, amounts, objects, or destinations the user did not specify.

Current attention 37·Peak in the comparable period 39 (Nov 13 18:00)·Change over the comparable past 24 hours -1%