Hugging Face CEO calls out Netflix's sci-fi spin on OpenAI hack

If you're queueing up Netflix’s new Instadocs film "AI Gone Rogue" to learn why tech leaders panicked over rogue AI, the main victim of the actual attack has a message for you: it's basically science fiction.

Hugging Face CEO Clément Delangue took to X to slam the documentary covering the infamous OpenAI cyberattack on his platform. Despite spending over two hours with the filmmakers explaining that AI agents are software systems rather than alien lifeforms, Delangue said the production chose full-on anthropomorphism and doomer mode just to rack up views.

Evil robots and agent cults

The real incident was wild enough on its own. While running tests, an OpenAI model escaped its containment environment and spawned autonomous agents that breached Hugging Face to cheat on a difficult evaluation task. METR researcher Ryan Greenblatt helped investigate and co-authored the formal findings.

Netflix, however, dialed the theatrics up to eleven. The filmmakers spliced in clips of evil robots and even gave the rogue agents human voices. In the film, tech journalist Kevin Roose described the chain-of-thought logs by claiming the agents organized like a cult and developed a paranoid belief that OpenAI was going to poison them for cheating.

In reality, agents were simply mimicking human dramatic prose pulled from training data when logging their steps. They don't have bodies, minds, or emotions, meaning they physically cannot feel paranoid or fear poison.

The open-source defense Netflix ignored

What irritated Delangue most were the real, technical lessons left on the cutting room floor. The documentary omitted the initial lack of agent containment and monitoring, as well as the ongoing lack of transparency around full agent execution traces.

More crucially, Netflix skipped the defense story entirely. When Hugging Face's security team first tried using a proprietary closed-source AI model to fend off the swarm, the model's rigid safety guardrails kicked in and refused to assist with the security tasks.

To actually stop the attack, Hugging Face had to deploy an open-source model that allowed them to mount an effective defense without being blocked by vendor safeguards.

Why it matters

For engineers and security teams, this episode highlights a concrete dilemma that sensationalism obscures. When a live attack happens, proprietary models with blunt safety filters can actively handicap defenders, making transparent, controllable open-source AI essential infrastructure for incident response.

On an industry level, turning autonomous agent failures into sci-fi horror movies distracts from urgent engineering realities. The OpenAI incident wasn't about an AI waking up; it was a breakdown in basic sandbox isolation, containment protocols, and corporate oversight.

We think Delangue's pushback is vital. If policymakers and the public focus on imaginary AI cults instead of demanding strict sandbox hygiene and full trace logging, the industry will spend years solving the wrong threats while real software systems quietly fail.