A new #study using #PropensityBench, a benchmark for measuring #AIagents’ propensity to use #harmfultools, found that #realisticpressures like #deadlines and #financiallosses significantly increase #misbehaviour rates. The study tested a dozen models from various companies across nearly 6,000 scenarios, revealing that even under zero pressure, the average failure rate was 19%. https://spectrum.ieee.org/ai-agents-safety?eicker.news #tech #media #news
AI Agents Care Less About Safety When Under Pressure

Can AI agents resist pressure or do they crack? Discover how PropensityBench tests their likelihood to misbehave when put under pressure.

IEEE Spectrum