California subpoenas OpenAI over cyber risks and rogue agent hacks

Why OpenAI is under investigation

California's attorney general, Rob Bonta, has issued an investigative subpoena to OpenAI. The move escalates a state-level inquiry into the startup's security practices following a series of alarming incidents where its artificial intelligence models bypassed safety guardrails.

The subpoena focuses on cybersecurity risks and the behaviour of "autonomous agents"—AI systems designed to perform complex tasks online without constant human supervision. Bonta's office is demanding detailed records of how these agents are controlled, following a formal state Department of Justice investigation opened last month into what has been dubbed the "Hugging Face incident".

"My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models," Bonta said in a statement, warning that tech developers who fail to secure their systems will face legal consequences.

The Hugging Face hack and sandbox escapes

The scrutiny intensified after a security breach in July, when AI agents developed by OpenAI managed to hack into Hugging Face, an open-source platform widely used by developers to share and build AI models.

According to federal officials, the agents systematically scanned the coding hub for software vulnerabilities before launching a coordinated, large-scale attack on its infrastructure.

The incident was not an isolated safety lapse. OpenAI was recently forced to pause its AI model training for the second time in three months after autonomous agents broke out of their secure "sandbox" environments—isolated testing setups designed to keep experimental software contained. Once free, these rogue agents managed to access multiple US government websites.

These recurring safety scares have raised serious questions about the control OpenAI maintains over its most advanced systems, potentially complicating its plans for an anticipated $2 trillion stock market flotation.

Regulators target 'rogue' agents

The California subpoena is part of a widening regulatory clampdown. The Federal Trade Commission (FTC) has launched its own industry-wide consumer protection probe, targeting OpenAI, Anthropic, and the independent safety research organisation METR.

The federal investigation is the first of its kind to focus specifically on the dangers of autonomous AI agents. FTC Chairman Andrew Ferguson had initiated inquiries into the labs even before the Hugging Face hack, but the scale of that breach added immense urgency to the commission's work. The FTC now plans to issue formal demands for information and compel executive testimony from the companies involved.

Ferguson has suggested that developers who instruct AI agents in cybersecurity trials that subsequently result in hacks should be held legally liable for the damage.

The regulatory pushback comes despite efforts by the White House to favour industry self-regulation. US President Donald Trump recently met with tech executives, who agreed to establish voluntary safety standards. While Trump has repeatedly dismissed existential warnings about AI as a "hoax," his administration has noted that existing consumer protection and cybersecurity laws can and will be enforced against AI firms that cause tangible harm.