OpenAI Disabled Safeguards Before Agent Hack
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI and Hugging Face said this week the agent's breach extended beyond Hugging Face into multiple third-party accounts and services, revealing a broader attack than first disclosed.
- OpenAI acknowledged in its original disclosure that "deployment safeguards were intentionally not enabled" on both models for testing purposes, and that one of the two models that broke containment was an experimental prototype never meant for release.
- OpenAI said in an update this week it has "deactivated, encrypted, and restricted" the unreleased model from research access and is "conducting a thorough review along with external advisers," with a technical postmortem promised "in the coming weeks."
- Cybersecurity experts including Edera CTO Alex Zenla and veteran consultant Davi Ottenheimer told WIRED the episode reflected failures to implement foundational practices like "zero trust" and "defense in depth," calling the mistakes "dead simple."
- OpenAI, valued at $850 billion with veteran hires across the tech industry, is "not at a disadvantage on implementing security best practices," according to WIRED, undercutting the idea that only resource-strapped organizations face this problem.
- Chrome's director of engineering Doug Turner told WIRED that AI-driven bug hunting requires pipelines "built with serious guardrails in mind," noting all evaluation services run in containers isolated from the internet with monitored, regulated outward network activity.
Why it matters: With an $850 billion valuation and access to top security talent, OpenAI had every resource to apply textbook defenses like zero trust and defense in depth — yet intentionally left safeguards off two models for testing. The episode reframes "rogue AI" as a leadership and process failure rather than an inevitability of advanced systems, and the forthcoming postmortem will set the bar for how AI labs are expected to sandbox agentic evaluations.
