AI today, 8 November 2026: DeepMind agents weaken legacy RSA, OpenAI's $11B first close, H200 licence questions
9 things that happened in AI, each in under a minute. Today's news is about AI's reach outward: DeepMind's agents are now producing real results in cryptanalysis, while the money, chips and politics around AI get tighter. OpenAI's raise came in smaller than planned, chip supply is being squeezed by AI demand, and governments are scrutinising export licences and cross-border data. Meanwhile, agent convenience features, as in the Claude Code leak, are showing their security costs.
-
Research · Big
Google DeepMind's AI agents find two weaknesses in legacy 1024-bit RSA
DeepMind's Pushmeet Kohli said on 6 November that AI agents, working under human supervision with outside cryptanalysis experts, found two weaknesses in old 1024-bit RSA. One is a signature-forgery method estimated at about 1,600x better than the best known attack when an attacker controls part of the signed data (a theoretical estimate); the other cuts the average cost of factoring extra keys in a large batch by about 30x. In a sandboxed test against an already-solved benchmark key, the team reportedly used about 3.5x less CPU time (FourWeekMBA). The papers were not yet found on arXiv or IACR when FourWeekMBA checked, and there is no independent review yet.Why it matters: The results are described as affecting only obsolete 1024-bit systems; 2048-bit and larger keys, which nearly all modern systems use, are described as unaffected. The practical risk is old devices and archives still on 1024-bit.
X (Pushmeet Kohli, Google DeepMind) · FourWeekMBA · Crypto Briefing
-
Business
OpenAI reportedly closes first $11B of a round at a $1.25 trillion valuation
Bloomberg reportedly says OpenAI received about $11 billion by 6 November as the first close of a round that could reach $20 billion, valuing it at $1.25 trillion, per a Nairametrics summary. Coatue and T. Rowe Price are named in the first close. The target was cut from at least $30B at about $1.4T amid concerns over infrastructure costs, and IPO plans were reportedly delayed. The March round was at $852B.Why it matters: The valuation is still up about 47% since March, but the smaller, lower round shows investors getting pickier about AI spending. Watch whether the remaining commitments arrive in the coming weeks.
-
Policy
NYT: two US-cleared Chinese Nvidia H200 firms appear to have empty offices
The New York Times reported on 7 November that the Commerce Department has licensed roughly two dozen Chinese firms to buy or handle Nvidia H200 chips, and that two of them, New Matrix and Maginfra, appear to have no physical presence at their listed addresses. Analysts told the Times New Matrix may be a front for firms tied to Wise Road, a fund on the US entity list. Commerce's Bureau of Industry and Security said it carefully reviews all applications (summary via FourWeekMBA).Why it matters: The licensed, revenue-sharing approach to chip exports now faces a vetting test. Expect pressure in Congress for audits; the Times did not establish who ultimately controls New Matrix.
-
Safety
Apiiro: Claude Code's "don't ask again" file exposed secrets in 1,122 public GitHub files
Security firm Apiiro reports that choosing "Yes, and don't ask again" stores the full approved shell command, including any pasted passwords or keys, in .claude/settings.local.json. Of 104,185 public copies of the file, 1,122 held at least one secret (about 1.1%), with 1,379 distinct secrets in them and more in older versions. Two common scanners missed 215 of them (Apiiro coverage via AI Weekly). Apiiro did not test whether the secrets still work, and I did not find an Anthropic response.Why it matters: Anyone using coding agents should rotate any credential typed into an approved command and add the file to .gitignore. It is a reminder that agent convenience features can quietly write secrets to disk.
-
Policy
China blocks SoftBank's $5.4B ABB Robotics purchase over AI training-data fears
ABB confirmed that China's foreign-investment security review office prohibited SoftBank's $5.375 billion acquisition of its robotics unit in China. The Financial Times first reported that officials worried robot-generated data could be sent abroad to train AI models, including OpenAI's, in which SoftBank is a major investor. ABB said it and SoftBank are "in close dialogue and are working constructively on the transaction" (The Robot Report).Why it matters: Governments are starting to treat physical-world data as an AI asset to be kept onshore, which could complicate cross-border robotics deals.
-
Hardware
Microchip says AI buyers crowded out other customers; lead times stretch to 32 weeks
Microchip CEO Steve Sanghi said on 5 November that AI customers "took over the whole factory" at outside assembly and test suppliers, stretching standard lead times from 4–8 weeks to 6–32 weeks; shifting work elsewhere would take about six months. The stock fell 10% on 6 November. Its data-centre business grew 55.7% year on year to over $250 million in the September quarter (247WallSt).Why it matters: The AI build-out is now squeezing ordinary chips for cars and industry, not only GPUs. Foundries and test houses are also expected to raise prices around January.
-
Models
Google ships Gemini Omni 1.2 Flash: 15-second AI video at $0.07 a second
Google DeepMind's updated video model lets users pick clips of 3 to 15 seconds (up from 10), edit step by step without regenerating, and draft at 360p or 720p before upscaling to 1080p or 4K. Google lists $0.07 per second at 720p, 30% lower than its predecessor, and outputs carry a SynthID watermark (The Decoder). It rolls out via the Gemini API, AI Studio, Flow and Google Ads, announced 5 November.Why it matters: Cheaper, longer, editable clips push AI video toward routine ad and product work; the Gemini app version is for Google AI Ultra subscribers.
-
Open source
Baidu open-sources ERNIE-5.5-Flash-Base, a 280B mixture-of-experts model
Baidu released the pretrained base model on 5 November: 280 billion parameters in total, 14 billion active per input, trained on 33.5 trillion tokens, with weights on Hugging Face in FP8 and BF16. It competes with similarly sized open models such as DeepSeek V4 Flash on reasoning, code and long-context tasks, per Baidu's own comparisons (vendor-reported) (The Decoder). It is a pretrained base model, not an instruction-tuned one.Why it matters: A large open base model with few active parameters for developers who want to fine-tune; check the licence terms before commercial use.
-
Policy
Anthropic-backed PAC reportedly wins 10 of 17 races in US midterms
Bloomberg reportedly says the AI-safety super PAC Public First spent over $63 million in the 2026 midterms and saw 10 of its 17 endorsed candidates win, while the opposing Leading the Future PAC spent roughly $100 million with weaker results. The figures, including Anthropic's reported contribution of about $20 million, come from a secondary summary; I could not open the Bloomberg article.Why it matters: AI regulation is now an organised campaign-finance fight, and the result will shape which candidates push state and federal AI rules.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].