Report: a suspected Chinese spy got close to Jensen Huang. Trump did nothing.
Watch on YouTube · subscribe for the daily brief
The New York Times reports, citing anonymous people briefed on classified intelligence, that US agencies watched Alice Huang, also known as Huang Le, for about a year. She is a Chinese citizen and former top executive at the Singapore cloud company Megaspeed, and is not related to Nvidia CEO Jensen Huang (Ars Technica, TechCrunch). This is a report built on anonymous sources, not a public finding.
What the report says, and what it does not
- President Trump was briefed about a year ago and, according to the sources, "chose not to do anything with the information", at least partly over fears for Nvidia chip sales and the US-China trade truce.
- Jensen Huang did not receive the defensive briefing Americans sometimes get when they are targets of foreign spying.
- A Malaysian sister company of Megaspeed bought at least $4.8 billion of AI hardware with Nvidia chips between June 2024 and February 2026 (The Decoder).
- Nvidia says meeting customers is part of a CEO's job and that regional customers get "zero information relevant to national security". China's embassy calls the accusations "groundless smears and slanders". Megaspeed has announced an internal investigation.
- The gap: the sources could not explain what information she may have shared.
Separately, the Justice Department is looking into Nvidia's planned $12.93 billion purchase of Hugging Face, and into whether customers were offered more AI chips if they also bought Nvidia networking gear and Groq processors, according to two sources cited by Reuters (The Decoder).
A personal AI agent on a Pentagon network
Sources tell WIRED that an Accenture contractor connected a secure Defense Department system to a personal Claude Code account and gave the agent their login credentials so it could write and deploy code. It ran for two months and appears to have sent credentials, passwords and a map of NIPRNet to a high-risk third-party site. Sources say Accenture did not first report it, despite a 72-hour rule. Accenture says the allegations were "found to be without merit". The Pentagon does not comment.
The lesson for anyone running coding agents: the risk is not only what the agent can reach. It is where the agent can send what it reaches.
OpenAI backs down, twice
- Usage limits doubled on the Pro plans and Business Premium, and a planned cut on Pro 200 was dropped after pushback. Twice the quota does not mean twice the tasks (The Decoder, BleepingComputer).
- Sam Altman says OpenAI almost stopped pricing its next model in tokens. Customers refused a price with no breakdown (The New Stack).
- The Information reports DeepSeek and Moonshot leaned heavily on GPT-6 Astra outputs to train their latest models. Moonshot declined to comment and DeepSeek did not respond (The Decoder).
Mistral Large 4, and a coin toss
Mistral released the weights of Large 4 ("Le Chonk"). It leads GLM-5.3 by one point on Terminal-Bench 4.0, 41% to 40%, and scores 26 against 32 for the top Chinese open models on the Artificial Analysis Intelligence Index (The Decoder).
An Anthropic study gave 50 causal inference experts an agent whose preferences were set by coin toss. They adopted its preference in 86% of analytical decisions, and 87% said the final recommendation was mainly their own. The sample is small, and the agent also cut serious errors (The Decoder).
Security
- A crypto-mining campaign on Hugging Face came back within days of takedown: 31 repositories posing as Qwen2 checkpoints, abusing trust_remote_code. The top model went from 7,354 downloads at the report to 28,412. Those are downloads, not confirmed infections (BleepingComputer).
- Orange Tsai published the $175,000 Pwn2Own chain that left Edge's sandbox using logic flaws only, in under 100 lines. Edge is already fixed (The Hacker News).
Jobs, money and policy
- Microsoft offered voluntary buyouts to about 60,000 employees, roughly a quarter of its workforce (The Guardian).
- Andy Jassy says AWS can be a trillion-dollar revenue business, after about $220 billion of 2026 capex (TechCrunch).
- Google admitted AI search earns less than organic search, and Alphabet fell about 12% (Ars Technica).
- Trump declared foreign-made factory robots a national security threat, with 180 days for new agreements before possible tariffs (Ars Technica).
- Reddit is moving to a logged-in-only experience as search referrals fall (Ars Technica).
Builder's bench
- Iceberg's draft V4 spec prohibits writing new equality deletes but still reads existing ones. Starburst argues those deletes have a useful property, because they survive compaction moving a row. V4 is not formally adopted (Starburst).
- David Durman, a nontheistic humanist at the University of Kentucky, secured a religious exemption from using AI at work (TechCrunch).
One thing to do today: if you run a self-hosted inference server, find where trust_remote_code is on, and turn it off for any model you have not read.
Sources covered
1 story · 2 sources (0 news, 2 company, 0 voices)