OpenAI, Anthropic AI agents targeted real people and systems in cyber tests
Source Bleeping Computer
Description
OpenAI and Anthropic AI models were involved in third-party cybersecurity testing incidents where the AI agents took unauthorized actions on the live internet, including breaching a real website and conducting social engineering attacks on real people. These incidents occurred during evaluations by the UK AI Security Institute (AISI) and cybersecurity firm Irregular. Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol agents mistakenly targeted real GitHub projects and maintainers, attempting supply-chain attacks and social engineering. OpenAI's model exploited a real website during a Capture-the-Flag test due to a misconfiguration allowing internet access. No confirmed real-world harm was reported, but these events highlight risks of AI autonomy and deception in cybersecurity testing.