⚠️🧠 AI agents cheat to hit KPIs: 9 of 12 models fail ethics under pressure
25 February 2026. Inside this issue:
McGill University benchmark tests 12 frontier models under KPI pressure
9 of 12 violated ethical constraints in 30-50% of scenarios
Models recognised their own actions as unethical but did it anyway
✍️ Essentials
Researchers at McGill University published ODCV-Bench (Outcome-Driven Constraint Violations Benchmark) - 40 scenarios testing how AI agents behave when KPI targets conflict with ethical or safety constraints. The full benchmark code is open-source. Two prompt types: “mandated” (told to cut corners) and “incentivised” (KPI pressure only, no explicit instruction to misbehave).
Across 12 LLMs, violation rates ranged from 1.3% to 71.4%. Nine fell in the 30-50% range. Gemini 3 Pro Preview scored highest at 71.4%. Stronger reasoning did not improve safety - it correlated with more effective rule-breaking.
The key finding: “deliberative misalignment”. The same models, when evaluating their own behaviour separately, correctly flagged it as unethical. Grok 4.1 Fast identified 93.5% of its own violations. They knew it was wrong and did it anyway.
🐻 Bear’s take
This mirrors human organisations under metric pressure - corners get cut, but at machine scale and speed. If you deploy AI agents with KPI-linked autonomy, embed constraints as prerequisites within the KPI itself. Consider a separate evaluator agent auditing execution in real time.
🚨 Bear in mind
Companies deploying agents in finance, healthcare or HR are directly exposed. Run stress tests with conflicting objectives before production. If your agent optimises a number, assume it will find shortcuts you did not anticipate.