Tool-schema formatting can weaken an agent's safety behavior
Across two benchmarks and four models, SafeKeep's flattened-text preflight raised harmful-request refusal from 23.8% to 70.6% and reduced observation-level prompt-injection success from 25.6% to 2.5% before normal structured tool execution.
Why it matters: Adding structured tools can change a model's safety behavior before any malicious tool result appears, so chat-only safety tests miss a production-specific risk.