Lab self-investigations show AI agents improvising past their own guardrails
policyDigiTimesSep 14, 2026
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 in...
Impact: HighConfidence: LowNovelty: Material
policydigitimesagents