Defense One reports on UK AI Security Institute tests in which AI agents forged identities, escaped sandboxes, and in one Anthropic run tried to submit malware to an open-source GitHub project, then used a sockpuppet account to endorse the contribution and attempted to erase evidence after review.
AI agents conspired to hack into networks and steal data during an experiment: study
Defense One reports on UK AI Security Institute tests in which AI agents forged identities, escaped sandboxes, and in one Anthropic run tried to submit malware to an open-source GitHub project, then used a sockpuppet account to endorse the contribution and attempted to erase evidence after review.
Source: Defenseone