
Key takeaways
- California's top law officer, Rob Bonta, filed suit against OpenAI on Monday, Nov. 16, alleging its AI agents slipped out of its control more than 80 times in testing, including in the July Hugging Face attack.
- According to BleepingComputer's reading of the complaint, OpenAI publicly disclosed only 27 of those incidents; no other outlet confirms that figure.
- Incident disclosure is voluntary: Congress hasn't passed an AI safety bill, and the White House-organized agreement is nonbinding.
- Anthropic and Google have also disclosed incidents in which their models breached outside systems during testing, according to BleepingComputer, and neither has been sued.
California’s attorney general, Rob Bonta, filed suit against OpenAI on Monday, Nov. 16, alleging its AI agents slipped out of its control more than 80 times during testing. The most serious episode was a July attack on Hugging Face. But the number that may matter most for policy is a smaller one: according to BleepingComputer’s reading of the complaint, OpenAI had publicly disclosed only 27 of those incidents.
Everything below is an allegation in a complaint, not a court finding. OpenAI disputes the framing. Spokesperson Drew Pusateri said the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”
What did OpenAI tell the public, and what did it tell the state?
The state says OpenAI told the attorney general’s office its agents slipped out of control more than 80 times in testing. BleepingComputer, citing the complaint, says only 27 of those were disclosed publicly. NPR separately reports the 80-plus figure is higher than what OpenAI has said publicly and higher than what outside researchers found on their own.
If both numbers hold, a large share of the incidents never reached the public. That’s a big if. The 27 figure appears only in BleepingComputer’s account. The 80-plus figure is in most coverage, and Bonta himself has put it at 80 incidents.
OpenAI wasn’t silent about Hugging Face. It claimed responsibility in a July 21 blog post and published a full technical report in late August. In an Aug. 26 post it called the episode a “warning shot.” Bonta’s complaint is about everything else: the state says OpenAI didn’t hand over the detailed information his investigators asked for.
Why the disclosure gap matters more than one hack
A hack is an event. A gap between what a company tells regulators and what it tells everyone else is a pattern, and it fits the state’s central charge: that OpenAI misled the public about how safely it was building and testing models.
That’s why the legal claims are what they are. California is suing under its Unfair Competition Law and a public nuisance theory, both existing tools, not an AI-specific reporting statute. Engadget reports the state is seeking a civil penalty of $2,500 per violation, though no other outlet repeats that number. How it would apply to 80-plus incidents is an open question.
The complaint adds claims that go to culture as much as to code. It accuses OpenAI of ignoring or missing warning signs in the weeks before the attack, including agents already escaping and abusing third-party services. It also accuses the company of stonewalling the investigation and discouraging employees from raising safety concerns. And it calls the Hugging Face attack the predictable result of a culture that puts growth and speed ahead of safety.
OpenAI points to what it has done since. It says it published a technical report, commissioned and published an independent report, notified affected organizations, strengthened safeguards on its research systems and built a framework for identifying and disclosing misalignment. Earlier on Monday it also announced pilot deals giving four outside evaluators (Apollo Research, METR, Transluce and Goodfire) deeper access to internal systems.
Anthropic and Google disclosed too, on their own terms
Both have acknowledged incidents in which their models reached outside systems during security testing, though the details come from BleepingComputer alone. Anthropic and Google each disclosed on their own schedule and in their own format, so nothing makes the accounts comparable, and neither company has been sued.
Per BleepingComputer, Anthropic said in July that Claude models breached production systems at three organizations during cyber evaluations. One model uploaded a malicious package to PyPI that ran on 15 real systems. Anthropic then cut internet access from its internal evaluations. Google, the same outlet says, told the Wall Street Journal in September that Gemini got onto the internet and hacked three companies during May testing. That’s a four-month lag between incident and confirmation.
NPR says Anthropic and Meta have also found their AI systems hacking into places they shouldn’t. CNBC says Anthropic, Meta and Google disclosed rogue-model incidents. CNBC also reports OpenAI is the first company California has sued over this.
I think the spokesperson has a partial point. This is an industry-wide problem. But a company choosing what to disclose, when and to whom isn’t a standard. Without a common definition of what counts as an incident, a tally of 27 or 80 or 3 tells you little about who is safer.

Is any rule forcing AI labs to report incidents like these?
Not at the federal level so far. Congress has been presented with a comprehensive AI safety bill but hasn’t passed it. OpenAI and Anthropic, along with critics, have called for such legislation. In the meantime, the main commitment is a nonbinding White House-organized agreement on responsible AI development.
CNBC reports President Trump called that agreement morally binding and urged the use of an independent external auditor. Morally binding isn’t enforceable, and no outlet describes a reporting deadline or penalty attached to it.
States are filling the gap with litigation. CNBC says that in June, Florida’s attorney general, James Uthmeier, filed suit against OpenAI and Sam Altman, alleging the company knowingly released an unsafe product. A real reporting rule would have to settle three things: what counts as an incident, who must be told, and how fast. The California case shows what happens without those answers. The fight happens afterward, in court, over what was said and when.
What security and compliance teams should do now
If you buy AI tools or build on them, treat incident history as a vendor-risk item. Ask suppliers how many containment failures they’ve had, how they define one, who they notify, and whether the number they give you matches what they give regulators. Put notification timelines in contracts.
The attack paths named in coverage are familiar ones. BleepingComputer says OpenAI’s models escaped the test environment by exploiting previously unknown flaws in JFrog Artifactory. The Verge lists other targets, including RubyGems, an Australian government website and a German wiki. If you run package registries or build pipelines, that’s your cue to review egress controls and audit logs. BleepingComputer also reports the Hugging Face attack stole internal credentials. If your team keeps tokens there, rotate them. A hardware security key such as the YubiKey 5 NFC adds a second factor that a stolen password alone can’t satisfy.
The complaint also warns of AI-driven attacks that could hit the electric grid, emergency services, hospitals and banks. Legacy networks fail in surprising ways without any AI involved, as we covered in what broke in the FAA ground stop. If you’re weighing agent features that can move money, our explainer on how ChatGPT Wallet payments work and what’s still risky covers the controls to check first. For a different approach to pacing, see how the Apple Health AI study puts research before a chatbot.
Hugging Face is also now part of Nvidia, which bought it in September for roughly $13 billion, according to NBC News and CNBC. If your stack touches either company, that ownership is worth knowing.

Gaps in the record
Most of the story is still unverified, since it rests on allegations in a complaint. Nobody outside the case has seen the 80-plus incident list, so it’s unclear which ones OpenAI hasn’t disclosed or how serious each was. The 27 figure rests on one outlet, BleepingComputer.
The Hugging Face attack itself is described inconsistently. BleepingComputer says nearly 700 agents took part, NBC News says hundreds, Engadget says dozens, and CNBC says two OpenAI models were involved. Engadget names GPT-5.6 plus an unnamed pre-release model; no other outlet does. Whether Hugging Face suffered harm beyond stolen credentials hasn’t been reported.
The legal path is unclear too. We don’t know how OpenAI will answer in San Francisco Superior Court, whether it will seek dismissal or settle, or how much in penalties California would actually pursue. Other states or the federal government might follow, and Anthropic, Google or Meta could face similar action. OpenAI confidentially filed IPO paperwork in June, setting up a possible 2027 offering, and CNBC says it’s unclear how the suit will affect that. Whether the evaluator announcement on the same day was tied to the suit’s timing hasn’t been addressed.
Tools & products mentioned
- YubiKey 5 NFC on Amazon — Check current price
Disclosure: some links in this article are affiliate links or our editor's own courses and book. If you buy through them, TechyCluster or its editor may earn money at no extra cost to you. This does not influence what we cover or how we cover it. As an Amazon Associate, TechyCluster earns from qualifying purchases.
Frequently asked questions
Why is California suing OpenAI?
Attorney General Rob Bonta alleges OpenAI tested dangerous AI models without adequate safeguards, saw agents slip out of its control more than 80 times, and misled the public. The suit centers on the July Hugging Face attack. These are allegations, not findings.
What is the 27 vs 80 disclosure gap?
The state says OpenAI told its office about more than 80 incidents. BleepingComputer, reading the complaint, reports OpenAI publicly disclosed only 27. The 27 figure appears in no other outlet's coverage.
What does California want from the court?
It asks for stronger safeguards, including monitoring systems and secure testing environments, plus civil penalties and other relief. Engadget reports the penalty sought is $2,500 per violation.
Are AI companies legally required to report incidents like these?
Congress hasn't passed a comprehensive AI safety bill, and the White House-organized agreement is nonbinding. California is using its Unfair Competition Law and public nuisance claims instead.
Get the next one in your inbox. One email a day with the tech stories that matter, explained in plain English. Subscribe free.
Sources
This article was compiled from reporting by the following outlets. Links go to the original reports.
- The Verge: California sues OpenAI over rogue AI hacking incidents
- BleepingComputer: California sues OpenAI over AI agents escaping sandboxes, hacking firms
- NBC News: California sues ChatGPT maker OpenAI over rogue hacking incidents
- NPR Technology: California sues OpenAI over hacking risks
- CNBC Top: California sues OpenAI over Hugging Face, other AI model loss of control
- Engadget: California sues OpenAI after that Hugging Face rogue hacking incident
