Happy Monday. This is Frontier, The Defiant's weekly recap of AI and frontier technology.
TOP NEWS THIS WEEK
Nvidia's board opens an investigation into CEO Jensen Huang
Meta prices Muse Ember level with Anthropic and OpenAI
Fireworks says its October breach exposed billing data and datasets
ALSO IN THIS ISSUE
Google adds two systems to its TPU chip lineup
Dropbox plans a $350 million deal for Tigris Data
NHTSA opens a probe of Waymo's school-bus fix
Shield AI flies multiple Air Force drones as a team
Markey asks Google for AI-free Search in schools
CHIPS / GOVERNANCE
Nvidia Board Opens Investigation Into CEO Jensen Huang
Nvidia's board opened an independent investigation on Nov. 8 into reports that CEO Jensen Huang was the target of a foreign intelligence effort. Independent directors will run it with outside counsel. The board said it has "no reason to believe that either national security or NVIDIA secrets were compromised." It also said it has confidence in Huang. Nvidia shares traded at $176.77 on Nov. 9. That is 26.1% below the Oct. 6 record close of $239.24.
Huang wrote to employees the same day: "I was unfaithful to my wife." The New York Times reported that Huang told the board about a relationship with Alice Huang. She is a business executive whom U.S. intelligence had identified as working for China. China's Foreign Ministry called the allegation "purely unsubstantiated."
Three House committee chairmen asked Huang on Nov. 5 for a closed-door briefing. They chair the Foreign Affairs, China and Intelligence committees. They want to hear about his interactions with Alice Huang and about Nvidia's protections against espionage. They proposed Nov. 18 and asked for a reply by Nov. 11.
Nvidia was already in a bear market before the board's statement. The stock closed at $186.64 on Nov. 6, 22.0% below the record. It lost 9.3% on Nov. 5 alone, by our own arithmetic. The iShares Semiconductor ETF holds U.S.-listed chip stocks. It fell 30.5% over the same month. On Nov. 9 the two diverged. At about 11:50 a.m. ET, Nvidia was down 5.5%. The fund was up 6.8%.
What to watch: Nov. 11, Huang's reply to the House chairmen; Nov. 17, Nvidia's third-quarter results; Nov. 18, the proposed briefing.
AI MODELS / PRICING
Meta Matches Anthropic and OpenAI on Small-Model Prices
Meta released Muse Ember on Nov. 9 at $0.10 per million input tokens and $0.50 per million output tokens, the same base prices as Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna. A token is the unit of text a model reads or writes. API customers pay per token. Ember is Meta's lightweight model for repeated agent subtasks. The three models now differ on two things: the prompt length at which a higher rate applies, and the input types they accept.
Muse Ember: base rate up to 200,000 input tokens. Above that, the whole request costs $0.30 and $1.50. Accepts text, images, audio, video and PDFs.
Claude Haiku 5.5: base rate up to 100,000 input tokens. Above that, $0.50 and $2.50. Accepts text and images.
GPT-6 Luna: base rate up to 272,000 input tokens. Above that, the whole request costs twice the input rate and 1.5 times the output rate. Accepts text and images.
Meta's benchmark table puts Ember first on business workflows. On AutomationBench-AA, Meta reports 37.8% for Ember, 20.2% for Luna and 13.9% for Haiku. Haiku leads on coding. FrontierCode 2.0-AA judges whether code changes could be merged. Haiku scores 34.9% there, Ember 27.2% and Luna 25.2%. Meta advises pairing Ember with a larger model. Muse Spark, at $1.25 and $4.25, plans the work. Ember runs the subtasks.
Anthropic cut prices at the top of its range the same day. Claude Fable 5.2 and Claude Mythos 5.2 cost $8 per million input tokens and $40 per million output tokens. That is 20% below Fable 5.1. Anthropic reports 49.1% for Fable 5.2 on FrontierCode 2.0. Its Opus 5.5 model scores 45.5% and costs $4 and $20.
Our take: Choose a Small Model by Prompt Length and Input Type
Below 100,000 input tokens the three models cost the same. Price differs only for longer prompts. A 150,000-token prompt costs $0.015 in input fees on Ember or Luna. It costs $0.075 on Haiku, by our own arithmetic from list prices. Audio and video inputs work only on Ember. Haiku has the highest coding score of the three in Meta's table.
SECURITY
Fireworks Breach Widens as Claude Code Files Leak Secrets
Fireworks AI said on Nov. 9 that an October breach exposed customer billing information and, for certain customers, a subset of incomplete model weights and datasets. Fireworks runs and trains open-source AI models for other companies. Its first bulletin in October had reported no detected access to customer models or datasets. Separately, security firm Apiiro found 1,379 secrets such as database passwords and cloud keys in public copies of a settings file written by Claude Code, Anthropic's coding tool. In both cases the exposed data sat in storage around an AI tool: training files at Fireworks, saved settings in Claude Code.
Fireworks traced the wider exposure to a security incident on Oct. 23 at a cloud provider that hosts its research machines. An unauthorized party used that access on Oct. 24 to reach Fireworks development machines and run an unauthorized workload. Fireworks said it detected the activity that day. The exposed billing information includes addresses, invoice and payment histories, account balances and credit usage. Payment card numbers were not affected because Fireworks does not store them. The company has disabled self-service dataset uploads and downloads. It reports no detected impact on its Inference service, Training SDK or Nexus product. It has not said how many accounts are affected.
How Claude Code Settings Files Keep Credentials
Claude Code saves an approval rule in a local settings file when a developer chooses "Yes, and don't ask again" for a command. When the saved rule records the exact command, it keeps any password or token typed into that command. Apiiro examined 104,185 of these files in public GitHub repositories on Nov. 1 to 3 and found secrets in 1,122 of them. It did not test whether the credentials still work. Boris Cherny, Anthropic's head of Claude Code, said on Hacker News that the team was "looking into what we can do to prevent this." He said it would prototype built-in secret scanning. Apiiro recommends checking both Claude Code settings files and their Git history, then replacing any exposed credentials.
OTHER STORIES WORTH YOUR TIME
Google Cloud announced TPU 8tx and TPU 8i Max on Nov. 9. Its eighth-generation AI accelerator family now has four systems. TPU 8tx adapts Google's training chip to keep data and model weights protected while a model runs. TPU 8i Max puts two 8i chips in one package for workloads with very long inputs. Google's product page lists all four systems as "coming soon."
Dropbox signed an agreement on Nov. 3 to buy Tigris Data for about $350 million in cash. It expects to close in the fourth quarter. Tigris provides object storage for application assets, model weights and media files. Dropbox also introduced DBX Agent Cloud. It starts as standalone storage compatible with Amazon's S3 service. Filesystem and content-processing features are planned next.
The National Highway Traffic Safety Administration opened an investigation on Nov. 6 that covers 2,895 Waymo vehicles. The agency said most incidents of cars passing stopped school buses came after Nov. 17, 2025, the date Waymo said it had finished a software fix. The investigation will assess whether that recall remedy was adequate. Waymo told TechCrunch it has "received no citations since March 2026."
Shield AI said on Nov. 9 that its Hivemind software flew multiple U.S. Air Force Collaborative Combat Aircraft as a coordinated team in an Oct. 26 test. These are uncrewed aircraft designed to fly alongside crewed fighter jets. Shield did not name the aircraft types or say how many flew. The Air Force said in October that it intends to have 500 combat-ready systems by 2032.
Skild AI released Steel on Nov. 9 under the Apache 2.0 license. Steel is a physics solver, built with Nvidia, that simulates different robots and tasks together on a GPU. Skild reports up to 13 times the peak simulation throughput of Nvidia's PhysX engine in mixed-task tests, with each engine measured at its best batch size. Across 20 single-task benchmarks the reported advantage is 3.8 times.
Sen. Ed Markey asked Google CEO Sundar Pichai on Nov. 9 to let schools turn off generative AI features on K-12 student accounts and keep Google Search. Google says schools will be able to enable Search and its Gemini chatbot separately. Its help pages describe AI Overviews, the AI summaries in search results, as a core feature that cannot be turned off. Markey asked for answers by Nov. 30.
Anthropic said on Nov. 9 that Claude agents found better solutions than the best listed ones for 72 of 203 open problems in MIPLIB, a library used to test optimization software. It also reported 45 proofs that a solution is optimal. Anthropic says it has submitted the solutions. MIPLIB's public records, last updated Sept. 24, do not yet confirm them.
Even more this week:
METR estimated that Anthropic is improving its models at about twice its earlier rate after 11 weeks of access to the company's internal data. METR, a nonprofit that evaluates AI systems, said Anthropic redacted most of the evidence it found most informative.
Thinking Machines Lab released Inkling 2 Small on Nov. 5 through its Tinker service. The company says it intends to release the model's weights under the Apache 2.0 license next week.
Google cut the price of its video model 30% to $0.07 per second with Gemini Omni 1.2 Flash, released Nov. 5. The model makes clips of up to 15 seconds, up from 10.
Anthropic began a U.S. beta of Personal Finance in Claude on its Pro, Max and Ultra plans. The feature reads bank and card data through Plaid. It cannot move money.
Frontier is produced by The Defiant. This briefing is for informational purposes only and does not constitute investment advice.