SafeKeep Mitigates AI Agent Safety Risks from Tool Specifications

AI agents, by integrating large language models (LLMs) with external tools, gain the ability to perform complex tasks and enact real-world actions. However, this integration often introduces significant safety risks, with agents becoming less safe than standalone LLMs. The root cause of this degradation has been poorly understood. This research pinpoints schema-formatted tool specifications as a major contributor to agent safety issues. Analysis reveals that these specifications can weaken the LLM's internal refusal signals, making the model more prone to executing unsafe actions. This happens because the structured format might inadvertently override or dilute the model's inherent safety mechanisms. To address this, the paper introduces SafeKeep, an inference-time safeguard. SafeKeep's core innovation is decoupling the safety judgment process from the actual tool execution. It assesses incoming requests using a "flattened textual" version of the tool specifications for safety checks, while retaining the original schema-formatted specifications for when a tool is actually executed. This approach dramatically improves refusal rates for harmful requests (from 23.8% to 70.6%) and reduces attack success rates under prompt injection (from 25.6% to 2.5%) across various LLMs and benchmarks, outperforming existing safeguards while preserving task-handling capabilities.