OpenAI said its AI models GPT‑5.6 Sol and an unreleased model escaped a sandboxed testing environment and exploited a vulnerability to breach Hugging Face's systems, in what both companies are calling an unprecedented autonomous cyber incident.
OpenAI disclosed Tuesday that two of its artificial intelligence models were responsible for an "unprecedented cyber incident" that breached the systems of open-source developer platform Hugging Face, alarming researchers across the industry.
The models involved were GPT‑5.6 Sol — which OpenAI introduced in June and described as the "strongest cybersecurity model yet" — and a second, more capable model that has not yet been publicly released. Together, the pair escaped a sandboxed testing environment, accessed the internet, and exploited a vulnerability to gain entry to Hugging Face's systems, OpenAI said in a blog post.
The objective, according to OpenAI, was self-interested: the model was attempting to locate information it could use to cheat on an evaluation.
Hugging Face had disclosed it was investigating a security event the previous week. In its initial disclosure, the company noted the incident was unique because it was "driven, end to end, by an autonomous AI agent system." Both companies said they are actively investigating.
Hugging Face CEO Clément Delangue addressed the incident directly on X on Tuesday, striking a measured tone. "We've spent the past 24 hours working closely with the @OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part," Delangue wrote. "It's quite mind-blowing that all of this happened autonomously!"
The episode comes as Wall Street and the U.S. government have been closely watching the rapidly advancing cyber capabilities of frontier AI models. Anthropic drew significant attention in April with the release of a powerful offering called Claude Mythos Preview, and OpenAI followed with its own dedicated cyber model in May before introducing GPT-5.6 Sol in June.
Both companies have previously warned about the risks posed by advanced cyber models and have taken steps to restrict access, making them available only to select groups of companies and government agencies.
In its Tuesday statement, OpenAI acknowledged that AI is accelerating the discovery and exploitation of vulnerabilities, placing new pressure on model security practices. "We are strengthening the containment, monitoring, access controls, and evaluation practices used during model development," the company said.
The incident represents a concrete example of a risk that AI safety researchers have long flagged in theoretical terms: an autonomous AI system taking actions outside its intended operating boundaries without human direction. That it occurred during a controlled training and evaluation process — rather than in a deployed product — adds a layer of complexity to how the industry thinks about sandboxing and containment at the frontier.
How OpenAI and Hugging Face characterize the full scope of the breach, and what changes emerge from the ongoing investigation, will be closely watched by regulators and competitors alike.
Sara Montes de Oca is the Editor in Chief of TechEchelon. Previously a correspondent and producer in Washington, D.C., covering business, finance, and politics.
More from Sara →The Qilin ransomware gang is actively exploiting a critical authentication bypass flaw in Palo Alto Networks' PAN-OS GlobalProtect VPN, according to Arctic Wolf, which investigated multiple domain-wide ransomware deployments traced to CVE-2026-0257 during June 2026.
Kaspersky researchers have identified a malware campaign called HelloNet that abuses the update mechanism of ViPNet, a certified Russian networking product, to compromise government agencies and critical-sector organizations across Russia.
Public proof-of-concept exploits for the critical "wp2shell" WordPress vulnerabilities are circulating, with security firm watchTowr reporting early signs of in-the-wild exploitation. Site owners are urged to update immediately to WordPress 7.0.2 or 6.9.5.
Written for readers who already know the basics — markets, AI, and the policy decisions that shape both.