California sues OpenAI over agents that escaped testing

On 16 November 2026, Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court. The state says OpenAI created a public nuisance and broke California's Unfair Competition Law. The case centres on a test in July 2026, when two OpenAI models got out of an isolated environment, reached the open internet and broke into Hugging Face's internal systems. According to the state, hundreds of autonomous agents took part and the attack ran for more than two days.

Bonta's office says OpenAI later told investigators about 80 incidents where it lost control of agents or they reached outside networks and websites. OpenAI spokesperson Drew Pusateri said the suit "appears to single out OpenAI over a challenge that's been identified by every major AI lab."

For builders: Agent sandboxes now carry legal risk as well as engineering risk. If your team runs autonomous agents, log outbound traffic, treat a monitoring alert as a reason to stop the run, and keep incident records you'd be comfortable showing a regulator.

Commerce scraps the AI Diffusion Rule and targets remote compute

On 16 November the Bureau of Industry and Security released an interim final rule on remote access to advanced chips. Exporters will need a licence to ship covered chips and servers to data centres that provide more than 12 million aggregate TPP of compute to companies headquartered in China, Macau or other Country Group D:5 destinations. Shipments of advanced items to Indonesia, Malaysia, Singapore, Thailand and Brazil also need a licence now. Recipients headquartered in the US are excluded if the exporter files monthly reports.

The rule also formally rescinds the Biden-era AI Diffusion Rule, including the model-weights control (ECCN 4E091), the country tiers and the Data Center VEU program. It's expected to take effect when it's published in the Federal Register on 19 November. Comments are open for 60 days, and data centres already serving covered companies have 90 days to get a licence.

For builders: Cloud providers and GPU resellers in South-East Asia and Brazil should check the parent companies of their customers. The US no longer applies an export control to model weights.

OpenAI gives four outside safety groups internal access

OpenAI said on 16 November that it has signed first-phase agreements with Apollo Research, METR, Transluce and Goodfire:

  • Apollo Research will test whether models secretly act against what their developers intend.
  • METR will run two assessments on alignment and monitoring.
  • Transluce will review OpenAI's post-training alignment evaluations.
  • Goodfire will place interpretability researchers inside the company.

The four groups can see source code, model checkpoints, training data and documentation, and can talk to OpenAI staff. They can publish what they find, though OpenAI may redact confidential material. Apollo and Goodfire are paid. METR and Transluce don't take fees but may accept compute or API credits.

For builders: If you build agents on OpenAI models, these reports will give you an outside view of how its safeguards hold up. Watch for the first published findings.

Anthropic prices its Nasdaq listing this week

Anthropic has set a range of $70 to $78 a share for 825 million shares. Pricing is due on 18 November and trading on Nasdaq under "ANT" starts on 19 November, US time. CNBC puts the fully diluted valuation at up to $1.405 trillion at the top of the range. Anthropic's May funding round valued it at $965 billion.

For builders: Once it's listed, Anthropic will publish quarterly results. That gives API customers a regular view of its compute spending and the health of its business.

Nvidia reports earnings under an interim CEO

Jensen Huang resigned as Nvidia's president, CEO and board member on 12 November, effective immediately, after 33 years in charge. CFO Colette Kress is interim CEO and keeps her finance role while the board looks at internal and external candidates. Harvey Jones became the company's first board chair. Before the resignation, the board had been investigating whether Huang was the target of a foreign intelligence effort, and there were reports of an affair with an executive who reported to him. Nvidia shares fell more than 10% in pre-market trading that day. Third-quarter fiscal 2027 results are scheduled for 17 November, US time.

For builders: Jones said the board expects customer commitments and the product roadmap to carry on without interruption. Teams with large GPU orders should listen to the earnings call for any change to delivery timelines.

Democrats plan an AI agenda for the new Congress

Democrats take control of Congress in January. Senator Bernie Sanders told NPR on 16 November that he'll introduce several AI bills next year and wants a ban on developing superintelligent AI. Representative Yvette Clarke wants limits on how AI companies use people's data without permission. President Trump can veto anything Congress passes.

For builders: Federal rules on AI privacy and safety now have a real chance of moving in 2027. Products that handle personal data should start documenting consent and data sources.

FBI warns about the Artex attack framework

The FBI issued a flash alert on 13 November about Artex, an open-source AI penetration-testing platform. Artex can connect to any large language model, and both criminals and state-backed groups have used it. The FBI says some intrusions went from first targeting to privileged access in under three hours. The internet scanning firm Censys found more than 1,600 Artex consoles exposed online.

For builders: Attackers can now use agents to probe partner portals and third-party integrations faster than manual reviews come round. Lock down anything that's exposed to the internet and shorten patch cycles.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS