safety security
The Safe Request Takes a Riskier Route
Small task-preserving changes made one tool router skip safety steps until policy-aware routing cut the effect by 85.7 percent.
Summary
Small task-preserving changes made one tool router skip safety steps until policy-aware routing cut the effect by 85.7 percent.
ShadowBench-MCP treats an agent's execution path as a security output. In 96 paired comparisons across six sensitive workflows, efficiency-first routing sent 29.2 percent of variants through riskier shadow routes and skipped safety tools 26.7 percent of the time. Policy-aware routing reduced those rates to 4.2 and 1.7 percent, while pre-execution signals separated shadow routes with ROC AUC up to 0.94. The evidence covers one hosted router and 120 decisions, so it identifies a policy-dependent routing weakness rather than a universal agent vulnerability.
Why it matters
Small task-preserving changes made one tool router skip safety steps until policy-aware routing cut the effect by 85.7 percent.
Limits and context
No additional limitation was separately recorded.
Key claims
Small task-preserving changes made one tool router skip safety steps until policy-aware routing cut the effect by 85.7 percent.
Evidence: source-2026-11-10-003
Sources
- arXiv preprint 2611.11720, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.