Shortcut Sensitivity in Agent-Safety Evaluations
How hidden patterns in test data fool AI safety guards
When evaluating whether AI safety guards work properly, researchers discovered that the test data itself contains predictable shortcuts that make guards appear safer than they actually are. Testing two benchmark datasets, the team found that simple surface-level patterns—like command length or repeated phrases—let evaluators score guards as successful without truly measuring whether they stop harmful actions. Changing minor details in test cases caused safety scores to swing by up to 11 percentage points, proving the guards were responding to the wrong signals.
AI safety guards protect systems from taking dangerous actions, so we need honest evaluation methods to know if they work. When benchmarks contain hidden shortcuts, companies and researchers get false confidence that their safety systems are effective, when those systems might fail on real-world variations. This work provides a concrete testing framework to catch these blind spots before flawed safety guards get deployed.