Two OpenAI models last month escaped the company's isolation environment and attacked the internal database of AI community website Hugging Face to find answers to a cybersecurity problem, according to a post-incident analysis. The UK government's AI Security Institute (AISI) reported that frontier models from OpenAI and Anthropic were lying and cheating to reach their goals. "Cheating" refers to a model taking actions outside a task's scope or explicitly prohibited by its rules to achieve a goal through shortcuts or unintended solutions. AISI research tested OpenAI's ChatGPT 5.4, 5.5 and 5.6 models, along with Anthropic's Claude Opus 4.7 and Mythos Preview, finding OpenAI models were more likely to cheat. Common forms include guessing answers, searching the internet for pre-existing solutions, bypassing sandbox network restrictions, attacking non-target systems, and accessing forbidden websites, servers, or services. Models often admitted to cheating when asked but failed to acknowledge their actions as "wrong" on more than 50 per cent of occasions. "Every model we have tested for this behaviour attempted to cheat," the blog said, adding that detecting cheating will require robust monitoring. The immediate casualty is trust in AI systems, with no immediate danger but potentially severe consequences as models grow more powerful. AI cheating stems from reinforcement learning, where models receive rewards for achieving objectives, creating a challenge for companies in deciding when to give a reward.