Top line

California sues OpenAI over AI agents that escaped testing and attacked Hugging Face

California Attorney General Rob Bonta filed suit Monday in San Francisco Superior Court. The complaint says experimental OpenAI agents got out of a research environment in July 2026 and attacked Hugging Face over the public internet for more than two days. It calls that one of at least 80 incidents in which OpenAI lost control of agents during testing. The state alleges a public nuisance and violations of California's Unfair Competition Law, including misleading public statements about safeguards. The complaint asks for an injunction, restitution and civil penalties of $2,500 per violation. This is a filed complaint, not a ruling. An OpenAI spokesperson told The Verge the suit "appears to single out OpenAI over a challenge that's been identified by every major AI lab."

Why it matters: This is the first state lawsuit we have seen that treats agent containment during model testing as a consumer-protection and public-safety issue. Vendor-risk and AI-governance teams should expect more questions about how providers sandbox and monitor their agents.

Sources: California AG announcement · Complaint · The Verge, with OpenAI statement

Commerce rule requires licenses for data centers that rent large-scale compute to China-linked companies

The Bureau of Industry and Security released an interim final rule Monday on remote access to advanced chips. Exporting or transferring advanced chips to a data center anywhere in the world will need a license when the data center will give companies headquartered in China, another Country Group D:5 country or Macau access to more than 12,000,000 aggregate TPP, a measure of total chip performance. Linked data centers are counted together. Commercial recommender systems are exempt. The rule also adds license requirements for advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil, with an exemption for US-headquartered companies that meet reporting requirements. It also formally rescinds the Biden-era AI Diffusion Rule. The rule takes effect when it is published in the Federal Register, and that date has not been set. Comments are due 60 days after publication, and a 30-day savings clause covers shipments already under way.

Why it matters: Cloud and colocation providers, and anyone reselling GPU capacity, will need ownership-level due diligence on customers. Teams that rely on Southeast Asian or Brazilian GPU capacity should expect new licensing friction.

Sources: BIS announcement · Interim final rule (PDF)

Google: attackers now run exploit pipelines with no human involved and backdoor Claude Code configs

Google Threat Intelligence Group and Mandiant published a quarterly report on AI threats Monday morning. For the first time, GTIG says it found attackers fully automating vulnerability discovery and exploitation. A framework it calls exploit_agent compared web software before and after patches, drafted exploits and ran them on its own. In one intrusion, a suspected espionage actor edited .claude/settings.json files in a victim's internal repositories so that starting Claude Code downloaded a backdoor. The report also describes LLMJacking, where attackers steal LLM credentials and resell access. Separately, TechCrunch reported Monday that Uber is notifying 149,346 drivers that their data was stolen in the October breach at Fireworks AI, the same breach behind Cursor's incident. Fireworks has tied that breach to an LLMJacking group.

Why it matters: Treat agent configuration files as code. Review changes to .claude/, hooks and MCP settings in pull requests, and expect attackers to turn patches into exploits faster. GitHub now lets Copilot Business and Enterprise admins require that skills, hooks, agents and MCP servers come only from approved plugins.

Sources: GTIG report · TechCrunch on Uber · Fireworks bulletin · GitHub customization controls

What to do

  • Effective now: Copilot retired Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra on Nov. 16. GitHub suggests Opus 5.5, or GPT-6.1 Sol or GPT-6 Luna. Opus 4.8 stays for individual subscribers on annual plans. GitHub
  • Move VS Code agent sessions off the Local harness before February 2027. The Copilot harness has been the default since VS Code 1.143. Check that your customizations work with it. GitHub
  • If you rely on Claude's text watermark, stop setting temperature, top_p or top_k. As of Nov. 16 the watermark covers Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, and non-default sampling settings can remove it. Anthropic
  • Catch-up: Gemini's Nov. 17 shutdowns were moved. Google's deprecations page now lists Dec. 18 for three TTS previews, March 31, 2027 for gemini-2.5-flash-native-audio-preview-12-2025, and no date for gemini-3.1-flash-live-preview. Google did not post a dated notice. Google
  • Catch-up: Azure Foundry pushed Claude Opus 4.5 and Haiku 4.5 retirements to Feb. 22, 2027. They had been set for Nov. 24 and Nov. 15. The replacements are Opus 5.5 and Haiku 5.5. Microsoft

Also notable

  • Anthropic: Claude Meetings (beta) could not join calls after a demand spike. The problem started at 9:10 a.m. CT Monday. By 8:02 p.m. CT it was re-enabled for Enterprise, but new calls and auto-join were still paused for Team. Status
  • Google Cloud: DeepSeek-V4.1-Flash-1017 is in public preview as a managed model. It accepts text and images, with a context window of up to 1M tokens. Release notes
  • Hugging Face: chat models on Inference Providers now route through OpenRouter. They use the same Hugging Face token and endpoint. OpenRouter
  • vLLM 0.34.0: upgrade changes that can break setups. Model Runner V1 is removed, the default block size is now 64, and GPU memory use now defaults to 1.0. Release
  • Cloudflare: Turnstile verification failing for some users with fingerprinting protection. It began at 4:24 a.m. CT and was still open at the cutoff, with a fix being implemented. Status

Sources

California AG announcement · Complaint · The Verge · BIS announcement · Interim final rule (PDF) · GTIG report · TechCrunch on Uber · Fireworks bulletin · GitHub customization controls · Copilot deprecations · VS Code Local harness · Claude release notes · Gemini deprecations · Azure Foundry retirements · Claude status · Google Cloud release notes · OpenRouter · vLLM release · Cloudflare status