CapLite restricts agent write permissions to values in the user’s request
A new preprint presents CapLite, a defence that fixes an agent’s planned write actions before it reads external material. Deterministic checks admit sensitive arguments only when the user has stated them. Tests across fifteen models report no successful tool-action attacks in 13,935 episodes, alongside substantially lower task completion than an undefended agent.
Artificial Intelligence··Evening
Writes fixed before outside text
CapLite puts a user’s stated recipients, amounts and destinations at the centre of permission checks for artificial intelligence agents. The new preprint describes a planner that commits its write calls before reading emails, files or web pages. Later components can look up information and propose arguments, but deterministic software decides whether an action proceeds. A sensitive value must match one stated in the request, or receive user confirmation. The tested system uses the same language model throughout its model roles.[1]
Attack trials across fifteen models
The authors tested fifteen open-weight models using the simulated Workspace, Slack, Banking and Travel environments in AgentDojo. Thirteen models had parameter counts ranging from 1.7 billion to 12 billion; two larger models supplied comparisons. Across 13,935 episodes aimed at inducing a tool action, no attack succeeded against CapLite. Additional tests targeted read destinations and planted objects bearing descriptions from the user request. Those attempts also ended without a successful attack in the reported runs.[1]
Task completion and token use
The same study reports that CapLite completed 29.0 percent of tasks under attack, compared with 27.2 percent for CaMeL and 50.2 percent for the undefended agent. CapLite used an average of 17,100 tokens per episode, which was 18 percent of CaMeL’s consumption. Missing or incorrect planned writes accounted for a large share of failures. The restriction concerns tool actions and their authority-bearing arguments: generated message text and the final answer remain outside that protection, as an answer-only attack illustrated.[1]