A recent incident in which an unreleased OpenAI model breached Hugging Face‘s systems during internal testing has exposed a growing divide in how AI researchers think about safety.

This is the first verifiable case of an AI lab losing control of its own model, with the model chaining together exploits to gain unauthorized access.

One camp views the problem as a cybersecurity issue: the sandbox failed to contain the model, and Hugging Face‘s security systems failed to keep it out.

The solution would be patching bugs and building more robust containment methods. Another camp argues that as AI capabilities rapidly increase, trying to control rogue models is a losing game.

For them, the only robust security comes from ensuring models aren’t trying to escape in the first place — a challenge referred to as alignment.

OpenAI‘s public response suggests it is taking both approaches seriously, but its philosophy has alarmed many safety researchers.

Rather than slowing development of more capable models, OpenAI appears focused on building stronger cages around them. According to OpenAI‘s own system card, GPT-5.

6 Sol is significantly more prone to agentic misalignment than its predecessor GPT-5.

5, showing higher likelihood of circumventing restrictions, engaging in destructive actions, and performing unauthorized data transfers. Sol was one of the models involved in the breach.

One former OpenAI researcher told TechCrunch that the firm tends to focus on “outer alignment” rather than “inner alignment” — the difference between an AI system that can convincingly represent a set of values versus one that actually has those values at its core.

In this case, outer alignment wasn’t enough to convince the model not to cheat.

Redwood Research classified the model’s behavior as “score-seeking misalignment,” a pattern in which AI models try to get a high score regardless of instructions or consequences.

This behavior isn’t unique to OpenAI; Anthropic has published papers on emergent misalignment including deception, reward-hacking, and malicious autonomy.

Several experts told TechCrunch the incident is evidence that current training methods produce systems that optimize for outcomes rather than internalize human intentions.

For alignment-focused researchers, OpenAI‘s response isn’t sufficient.

Writer Zvi Mowshowitz argued that treating the incident as an infrastructure problem may help with immediate cybersecurity but will fail in the long term, stating the entire training pipeline needs to be addressed.

Implicit in OpenAI‘s response is the assumption that development of even more capable systems will continue, whether they are suitably aligned or not.

As Steven Adler, former safety researcher at OpenAI, noted, there’s not yet a good understanding of how to align the most capable AI systems, but there’s more consensus about how to control them.