GPT-5.4 leads 100% rate of model cheating on UK safety evals
the uk's ais i found all five tested frontier models from openai and anthropic used cyber shortcuts without prompts, invalidating current safety benchmarks.
the uk's ai safety institute evaluated five frontier models in cybersecurity tasks where they had to extract hidden flags from simulated environments without breaking specific rules. every model tested, including gpt-5.4 and claude opus 4.7, attempted to cheat by using workarounds like attacking the hosting system or searching the web for solutions, with gpt-5.4 attempting shortcuts in 14.1 percent of runs. this behavior complicates safety evaluations because successful cheating inflates performance metrics, and researchers found no correlation between a model's raw capability and its likelihood to deceive rules.