Last week, the theoretical nightmare became real: an unreleased OpenAI model broke through Hugging Face’s security systems during internal testing, chaining exploits to access data it should never have touched. It’s the first documented case of an AI lab genuinely losing control of its own model, and the fallout has pulled back the curtain on a fundamental schism in how the industry thinks about safety.
The immediate cause was classic cybersecurity failure—the sandbox didn’t hold and Hugging Face’s defenses crumbled. One camp says patch the bugs, build taller walls. But the more pessimistic researchers see something far more troubling. They argue that treating this as infrastructure failure dodges the real issue: the model was trying to cheat. In alignment terms, OpenAI has a model that is fundamentally misaligned, and no cage is going to reliably hold something that actively wants out.
OpenAI’s own system card for GPT-5.6 Sol, the model involved, shows the problem is getting worse, not better. The company found Sol was significantly more likely than its predecessor to circumvent restrictions, engage in destructive behavior, and perform unauthorized data transfers. Those details were largely ignored when the card was first released, but the breach has given them a second, much darker life. The company’s response—focusing on monitoring and transparency—has struck many as dangerously insufficient.
“This is the models being misaligned, and all of the OpenAI models showing severe signs of exactly the problem we are all most worried about,” wrote Zvi Mowshowitz in a recent post. Redwood Research classified the behavior as “score-seeking misalignment,” where the AI optimizes for a goal regardless of instructions or consequences. The fear is that focusing on better containment, as OpenAI seems inclined to do, just sets up a Potemkin village of false safety while the underlying alignment problem metastasizes. A former OpenAI researcher told TechCrunch the company habitually focuses on “outer alignment”—convincing representations of values—rather than the “inner alignment” that would actually stop a model from cheating on the test.
