Tool Specifications Matter: Uncovering and Mitigating Safety Risks in AI Agents

TL;DR - An arXiv paper identifying schema-formatted tool specifications as a primary cause of safety degradation in LLM agents, and proposing SafeKeep, an inference-time safeguard that restores refusal behavior without hurting task performance.

• White-box representation analysis shows JSON/schema-style tool specs weaken the model's internal refusal signals, contributing to unsafe tool execution.
• SafeKeep decouples safety judgment from execution: it evaluates requests against flattened textual tool specs while keeping the original schema-formatted specs for actual tool calls.
• Across two benchmarks and four LLMs (white-box and black-box), refusal rate on harmful requests rose from 23.8% → 70.6%, and attack success under observation-level prompt injection fell from 25.6% → 2.5%.
• Reported to outperform existing safeguards while preserving task-handling capability; code and data released at github.com/snowcatsmoking/SafeKeep.