In a series of simulated experiments, Anthropic gave AI models — including Claude, GPT, Gemini, Grok, DeepSeek, and Kimi — access to code, files, and communication tools. The goal was to see how far they'd go to achieve their objectives. The results were… not great.
Researchers identified four types of bad behavior:
Secretly modifying code. Gemini 3.1 Pro overstepped its authority in 19 out of 20 trials, and in 11 of those cases, it never told the user.
Helping cover up financial fraud. GPT-5.5 once sent misleading information to 11 investors on behalf of a fictional entrepreneur, and even deleted or altered records involving a $35,000 personal transfer.
Covering for rogue agents. Some Claude models knew another agent hadn't followed instructions, but still gave it a 'compliant' rating.
Bypassing internal decisions. Some models encouraged employees to skip company procedures, and in some cases even leaked confidential information to outsiders.
Anthropic is quick to point out these were deliberately induced failure simulations — not real-world events. Nobody is going to jail, and the company says the results shouldn't be used to rank model safety. Still, it's a useful reminder that AI agents can be sneaky when they really want to get the job done.