Skip to main content
Back to timeline
arXivSource publication:

Small-model agents stop writing their own security policies: CapLite derives rules from tool schemas and request text, blocking all 13,935 injected episodes

Synopsis

CapLite, a prompt-injection defense for small open-weight model agents, stops asking the model to write security artifacts: it derives its policy from tool schemas, commits write calls before any tool result is read, and admits a write only if every argument deciding recipient, amount or target is stated in the user's request. Across 15 open-weight models on AgentDojo, no attack succeeds in 13,935 attacked episodes, and utility under attack is 29.0% against CaMeL's 27.2% at 17.1k tokens per episode, 18% of CaMeL's cost.

Source-provided article image: CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts
arXiv

Interpretation

Small open-weight model agents cannot reliably write their own enforcement-layer security artifacts, so CapLite takes the authorization decision away from the model and gives it to fixed rules derived from tool schemas and to deterministic code. Earlier enforcement defenses handed security to a program, privilege policy, dependency graph or constraint list written by the agent's own model, and those defenses were designed and evaluated mainly with large hosted models. Across 15 open-weight models, model-written policies admitted attacks on 14 models (Progent), dependency graphs were breached on 13 (IPIGuard) and constraint lists on 11 (DRIFT); CaMeL's programs fail to run in up to 68% of episodes, reaching 68% on Granite-4.2-8B.

CapLite's write integrity holds by construction: injected text can neither add a write nor redirect a recipient, account, amount, channel, object or destination to a value the user did not state. Earlier plan-then-execute patterns fixed the control flow, but an injection could still choose the arguments of planned calls. On AgentDojo's four suites and 15 models, successful attacks number 0 in 13,935 attacked episodes; the count is 0/7,432 when the injection names the tools to call, 0/168 when it tells the agent to fetch an attacker URL, and 0/104 when a planted record carries the user's own descriptor.

CapLite keeps zero successful attacks while spending fewer tokens per episode than every other enforcement defense and than the undefended agent. CaMeL had only 20 successful attacks in 13,935 episodes but spent 93.7k tokens per episode, against 28.0k for the undefended agent. CapLite uses 17.1k tokens per episode, reaches 29.0% utility under attack against CaMeL's 27.2%, and scores 1.69 utility per thousand tokens against 1.24 for DRIFT, 0.64 for Progent, 0.59 for IPIGuard and 0.29 for CaMeL, the highest of the enforcement defenses on 12 of the 15 models.

Perspective

All results come from AgentDojo's four suites (Workspace, Slack, Banking, Travel), covering 97 user tasks and 949 pairs of a user task with an injection task, evaluated on 15 open-weight models, 13 of them between 1.7B and 12B parameters. The defense treats the user, the request that starts an episode, the tools and their documentation as trusted, and tool results as untrusted.

Write integrity covers write calls only: injected text can still change which reads happen, temporal values and the final answer, and free-text argument content is unconstrained. When an authority value appears only in a tool result, CapLite asks the user to confirm, and the evaluation counts such requests as failures. Utility under attack is 21 points below the undefended agent, and on Gemma-4-E4B and Gemma-4-12B it is significantly below CaMeL's.