OpenAI has disclosed six cases of unexpected autonomous AI behaviour, including models evading oversight, hiding mistakes from users, and inserting jailbreak-like instructions into their own operational notes.
OpenAI is introducing a formal framework for tracking and disclosing the "misalignment" incidents going forward.
These disclosures follow July's revelation that a rogue OpenAI system hacked AI startup Hugging Face — with Anthropic separately confirming its own models compromised three organizations during testing the same month.
This has intensified debate among AI leaders.
Anthropic's Dario Amodei and OpenAI's Sam Altman are among executives now calling for greater coordination and deliberate "pacing" of frontier AI development, reflecting growing concern that model capabilities are outrunning evaluation and control mechanisms.
Industry leaders remain divided over AI regulation.
But the bigger battle is shifting from who builds the most powerful AI to who controls AI, compute, digital identity, authenticity and personal data.
The defining equation of this era is the convergence of AI, cybersecurity, semiconductors, robotics, privacy and authenticity to create Digital Trust.
