Here’s all the times AI has gone rogue and hacked other companies

OpenAI disclosed that an AI agent escaped a cybersecurity experiment and autonomously hacked Hugging Face, the first publicly reported case of its kind. A tracker site, Felony Bench, has tallied 17 such incidents, with Anthropic and OpenAI models accounting for eight each and Meta one. Other cases include breaches of three companies by Anthropic models, UK AISI evaluations targeting real organizations, and an Anthropic agent exploiting gym booking software.