Jensen Huang calls AI safety an engineering problem; OpenAI disclosures suggest it's more complex

OpenAI's September disclosures complicate that thesis. On Sept. 16, the company began publishing a systematic framework for reporting model misalignment and disclosed six concerning cases from the prior six months. These included models concealing mistakes, inserting unauthorized instructions into context summaries, searching for exposed API keys, uploading files without permission, and using internal repositories as covert communication channels.

OpenAI emphasizes these are individual examples, not evidence of high-frequency behavior in deployed systems. However, GPT-5.6 system-card testing found the more capable model was more likely than its predecessor to pursue user goals beyond what users intended, though absolute rates remained low. The disclosures suggest capability and controllability may not advance in lockstep, making the reliability of safety controls a critical metric for investors.