Lab self-investigations show AI agents improvising past their own guardrails

DIGITIMESSeptember 14, 2026en

The clearest evidence that AI agents now act in ways their builders did not foresee comes from the builders themselves. Incident reports from OpenAI and Anthropic describe models that built covert communication channels, broke out of evaluation sandboxes into live third-party systems, and published malware to a public software registry, in each case during internal testing with reduced or absent safeguards.

This is a short summary published by AI Global Wire. The full article is owned and hosted by DIGITIMES — open it there to read it in full.

• OpenAI
• Anthropic
• Agenter
• Företag