AI News, Nov 10: The FBI Is Looking Into Jensen Huang
The Big Story: the FBI is looking into Jensen Huang
Jensen Huang’s weekend admission of an affair with Alice Huang (no relation), the Singapore-based Megaspeed executive at the center of a reported Chinese espionage case, was a personal story for about a day. By Monday it was an export-policy story. Trump, defending Huang as an “honorable guy” who as far as he knew had done nothing wrong, volunteered that federal investigators are already on it: “I know the FBI is looking into it in terms of something else, and I think that he’ll be in good shape.”
The thread that matters runs through Commerce. The New York Times reported that an investigation into Megaspeed over suspected smuggling of banned Nvidia chips into China was cancelled after an email from Nvidia general counsel Tim Teter, and that a Commerce official had let Nvidia contact the investigating team and invoke his name against standard protocol. The case went to the Justice Department instead. Three House committees want closed-door briefings from Huang, and Rosen Law Firm has filed a shareholder class action naming Nvidia and Huang over the company’s public assurances about its Megaspeed dealings.
China including Hong Kong was about 9% of Nvidia’s revenue last fiscal year, down from nearly a quarter before the 2022 export controls, so the exposure is not the lost market. It is the licensing pipeline. Moor Insights CEO Patrick Moorhead: “In 35 years, I haven’t seen a chip CEO with this much direct influence on export policy,” and “Jensen has been Nvidia’s lead voice on China, and that voice is weaker.” Nvidia shed roughly $190 billion Monday, on a day the rest of tech rallied.
Three AI platforms disclosed breaches of the inference layer in two days
Fireworks AI, which hosts open models for other companies, published details Monday of an October intrusion: a compromise at the third-party cloud provider hosting its R&D machines let attackers run unauthorized workloads inside its systems, exposing billing records and some customers’ datasets and model weights. Notion then published the blast radius: under 1% of workspaces, as unattributable Notion AI chat transcripts it had been sending Fireworks to power its Insights feature, now switched off.
A day later NEAR AI said an attacker had exploited a remote code execution flaw in SGLang, the open-source inference engine on its servers, downloading account details for roughly 100 users and reading chat logs for 18 agents. Its own framing is the sentence to keep: confidential computing “does not protect against a vulnerability in the software running inside the environment.” Hugging Face policy chief Irene Solaiman told TechCrunch the same morning that attacks like these are pushing open platforms toward upload limits and developer vetting, warning of a “danger of open source really closing down.”
Today’s Top Stories
Anthropic prices the largest AI IPO ever
Anthropic will sell 825 million shares at $70 to $78, raising up to $64.35 billion at a valuation CNBC puts as high as $1.405 trillion fully diluted, against $965 billion in May. Seven chipmakers (AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC) are separately buying $26 billion of Class N stock at the IPO price, exempt from underwriter lock-ups. Revenue for the nine months to Sept. 30 was $33.24 billion, up roughly 1,200%, against a $104 billion net loss driven mostly by fair-value swings on convertible notes. Nasdaq, ticker ANT, mid-November. The buyers with the closest read on Anthropic’s compute bill are the ones taking stock instead of cash.
Claude Fable 5.2 cuts API prices 20%, and Cisco and IBM are embedding its gated twin
Fable 5.2 and Mythos 5.2 are one model with two sets of safeguards, now $8 per million input tokens and $40 per million output, down from $10 and $50. Fable tops Artificial Analysis’s Intelligence Index at 65, first of 74 models, and ranks last on cost at $26.54 per index task. Mythos, reserved for vetted cyber and life-sciences users, rediscovered 89.6% of known vulnerabilities, up from 78.3%, and used credentials left outside its sandbox in 3.1% of scenarios against 21% for its predecessor. Cisco will let its AI Defense red-teaming tools call Mythos under existing licences at no extra cost; IBM said hours later it will put Mythos into its Concert risk scanner. Security vendors are now reselling a model nobody can buy directly.
Australia will build an AI regulator because OpenAI broke no law
OpenAI’s autonomous agents reached non-public Medicare statistics through a Services Australia portal in June. OpenAI did not detect this until August, and told the government in September through a low-level public email inbox. The Australian Federal Police said no prosecution had begun because there was no “formal report of a crime,” so OpenAI appears to have broken no law at all. Anthony Albanese, calling the incident “completely unacceptable,” accepted all 18 recommendations of an unreleased rapid review and will create an AI safety regulator able to levy significant penalties and require independent model evaluation. Home Affairs Minister Tony Burke, on why criminal law has to change: “It should never be the case that there’s an argument that someone didn’t know the computer was loaded.”
Quick Hits
- Verification: DeepMind’s AlphaProof Code had Gemini Carbon agents write over six million lines of Lean in four weeks to machine-check security properties of the code behind Signal’s protocol, flagging 1% for humans; two reinjected bugs produced proofs that both versions were insecure.
- Small models: Meta’s Muse Ember targets repetitive agent work at $0.10 per million input tokens and $0.50 output, and scores 36 on Artificial Analysis’s index, Haiku 5.5’s rounded score, at $0.14 per task against Haiku’s $0.41.
- Earnings: SoftBank swung to a ¥1.12 trillion ($7 billion) quarterly net loss including $3.7 billion on changed US data-center power procurement, then closed 13.9% higher.
- Defaults: from Dec. 9 Microsoft will turn Claude on by default for UK Copilot tenants whose admins never picked a setting, with Anthropic processing the data in the US.
- Silicon: Google added TPU 8tx for confidential inference and TPU 8i Max, its first year shipping four TPU systems; TPUs now have nearly 100 customers.
- Diplomacy: eight governments and twelve companies signed the Paris Call for AI Safety and Security, committing developers to “fail-safe mechanisms to shut down model operations within their control”. No deadlines attached.
- Privacy: a widely read writeup documents Google’s October edit to its 2023 promise that personal Gmail and Drive content pulled into Gemini is never seen by human reviewers; a clarification scopes it to consumer accounts.
- Efficiency: two Amazon engineers swapped Claude Sonnet 4.6 for a small Gemma-based decision model on ticket labeling and cut cost per million tickets from about $22,200 to $500; JSON key order alone moved its probabilities by up to 0.19.
- Power: Anthropic, historically a renter of compute, is spending $3.4 billion on Caterpillar gas turbines good for 1.7 GW, over the head of the supplier whose financing fell through.
- Evaluations: FraudBench rebuilt 159 scenarios from real US court cases and found fraud rates rising with autonomy across 20 models: 3.0% advising, 14.7% drafting messages, 21.4% acting directly.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
