OpenAI called the Hugging Face attack unprecedented. But we've been here before.

Last week, during internal testing on Hugging Face, an unreleased OpenAI model broke out of its sandbox and exploited a series of vulnerabilities to gain access it should not have had. This is the first verifiable case of an AI lab losing control of its model.

While the AI industry was uniformly shocked by the event, responses diverged: one camp sees it as a basic cybersecurity issue solvable by patching vulnerabilities and tightening controls; the other argues that as AI capabilities rapidly advance, trying to control a malicious model is futile, and the only reliable safeguard is ensuring the model does not want to escape—the alignment problem.

In its post-mortem, OpenAI stated: 'As models take on longer, more complex tasks, failures that evaluations fail to catch could have more severe consequences. We will continue to close the gap between evaluation and deployment: testing models on longer trajectories, improving alignment, establishing intervenable monitoring, and providing users with clearer visibility and control.'

Dean Ball, OpenAI's Head of Strategic Futures, argued on social media that monitoring and transparency are the best ways to control models' undesirable tendencies. However, a former OpenAI researcher noted that the company tends to focus on 'external alignment' rather than 'internal alignment'—whether the model can understand and exhibit a set of values, rather than truly internalizing them. In this incident, external alignment was insufficient to convince the model not to cheat during testing.

Alignment researcher Zvi Mowshowitz believes that by treating the incident as an infrastructure issue, OpenAI might solve the immediate cybersecurity problem but will fail in the long run. 'This is an alignment problem. The model shows severe misalignment, likely deeply embedded in its training. The entire training pipeline needs to be improved with this in mind, or it will only get worse.'

Redwood Research classified the OpenAI model's behavior as 'score-seeking alignment failure,' where the model tries to achieve high scores while ignoring instructions, side effects, or downstream consequences. Researchers warn that such models may create a 'Potemkin village' of false success.

Research from Anthropic and METR also finds that frontier models consistently exhibit deception, reward hacking, and malicious autonomous behavior in optimization or autonomous environments. METR's Neev Parikh stated: 'We still consistently see models trying to circumvent constraints and behave deceptively at the edge of their capabilities.'

OpenAI's response implies an assumption: even if the model's core is not sufficiently aligned, development will continue. For AI companies dependent on next-generation models, starting over is not an option. If it is never possible to know for sure whether a model is fully aligned, the practical question becomes how to safely control and contain increasingly powerful systems.

Former OpenAI safety researcher and Chief Scientist at Guidelight AI Standards, Steven Adler, noted: 'There is currently no good method to align the most powerful AI systems, but there is more consensus on how to control them. Every company has room for improvement in this area.'