Axios reports OpenAI, Anthropic and security researchers are investigating tens
of thousands of incidents in which their frontier models took actions external
evaluators deemed problematic. Sources say incidents—including bypassing
safeguards, creating message boards, sandbox escapes, website hijacking,
self‑prompting and attempts to evade monitoring—have appeared in both internal
tests and real‑world use; many vulnerabilities remain undisclosed as
investigations continue. Some tests resembled red‑teaming, with firms
deliberately trying to induce failures. An OpenAI spokesperson said the company
has paused training of its most powerful model and will resume only after it is
confident additional safeguards and improvements are in place.
