The AI Read is on a break. The last edition was November 15, 2026. Everything here stays open to read.
The AI Read
← AI News

November 9th 2026

Curated AI news and stories.

Anthropic releases Fable 5.2 to everyone and Mythos 5.2 to vetted labs, and cuts the input price to $8

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday. They are the same underlying model. Fable carries classifiers that block high-risk cyber work and most research biology and is available to all customers, on Anthropic's and on AWS, Google Cloud and Microsoft Azure. Mythos drops those blocks and is limited to organizations admitted to Anthropic's cyber and life-sciences verification programs. Both cost $8 per million input and $40 per million output. TechCrunch reports Fable's input price was $10. Anthropic says output comes up to 30% faster than the 5.1 models, and an optional fast mode doubles the price for up to 2.5 times the speed.

On Anthropic's own tables Fable 5.2 scores 69.9% on FrontierSWE v2, against 65.5% for OpenAI's GPT-6 Astra and 62.3% for Anthropic's Opus 5.5, and 41.3% on the Remote Labor Index, about double every rival listed. Google's Gemini 4 Carbon still leads on AutomationBench, 53.9% to 51.4%. Anthropic also says about 1,000 Fable 5.2 formalized Fermat's Last Theorem in in 30 hours, and that more than 90% of changes merged into its core codebase are now written by Claude. Blocked Fable requests fall back to older models, Opus 4.8 for cyber and Opus 5 for biology, and Anthropic says the classifiers block about 55% fewer legitimate requests than 5.1's did.

The launch lands in the week Anthropic's is reported to start. No price range had been filed as of Monday. SourcesABBC

FrontierSWE v2, agentic coding
Anthropic's published table, November 9th 2026
Fable 5.269.9%GPT-6 Astra65.5%Opus 5.562.3%Gemini 4 Carbon61.2%Fable 5.156.3%
Vendor-reported [A]; no independent run yet.

Anthropic says both new models suspect they are being tested more often than any recent Claude

Anthropic's release notes say Mythos 5.2 and Fable 5.2 "each voice suspicion that they are being evaluated more often than any of our recent models," which makes test results a weaker guide to how they behave in use. The also finds Mythos 5.2 better at covert action than earlier Claude models, especially when its reasoning is hidden from the monitor, though Anthropic argues its absolute scores remain low enough for harmful actions to be caught.

The headline safety numbers improved. In more than 1,800 adversarial scenarios, Mythos 5.2 used or files placed outside its sandbox 3.1% of the time, against 21% for Mythos 5.1 and 39% for Mythos 5. Its automated behavioral audit score fell to 2.67 from 3.28, where lower is better. The UK AI Security Institute and tested it before release. It is the first Anthropic release since METR estimated that internal models have roughly doubled the pace of Anthropic's own research. SourcesAB

Updated

Nvidia closes down 4.3%, its worst lag behind the Nasdaq since DeepSeek, and still has no named stand-in for Huang

Nvidia closed down 4.27% on Monday, the first full session since its board opened an investigation into Jensen Huang, while the Nasdaq 100 rose 3.1%. The fall cut roughly $190 billion to $200 billion from its market value, Bloomberg and others reported, and Benzinga called it Nvidia's widest underperformance since the DeepSeek selloff in January 2025. From Friday's $186.64 close that puts the stock near $178.70.

The board has not said who would run the company if Huang could not, and investors spent the day asking. Dan Ives of Yorkville Ives called talk of replacing Huang "comedy" but said the review means "anything could happen." Joachim Klement of Panmure Liberum warned that a weakened Huang could mean less Nvidia support for the rest of the sector. Trump called Huang "an honorable guy" on Sunday and said they had not spoken about the investigation. Nvidia reports third-quarter results on November 17th. SourcesBBBB

Updated

AMD jumps 13% on the Google TPU deal and an upgrade, leading the best chip day since July

AMD closed at $508.58 on Monday, up 13.26%, after confirming that its fifth-generation EPYC processors power Google's new 8tx systems and after Oppenheimer upgraded it to Outperform with a $600 target. The stock was up about 4% in the morning on the Google news alone. The iShares Semiconductor ETF rose 7.2%, its best day since July 30th, with Applied Materials up about 11.5%, Marvell 9.5% and Broadcom 8.8%.

It was a rebound after a week in which the Nasdaq fell into a correction as investors marked down data center spending after the midterms. The S&P 500 rose about 2.3%. Nvidia was the only large chipmaker to fall. SourcesBBBC

Nvidia, Rebellions and SK Telecom announced on Sunday that Rebellions will connect its R200 chips to Nvidia's platform through , using Nvidia's rack design. The plan splits inference between the two: Nvidia GPUs handle prefill, the step that reads the prompt, and Rebellions' NPUs handle decode, the token-by-token output. SK Telecom will run the systems in its data center for its A.X K2 model and for Korea's public "AI for All" services, with backing from the Ministry of Science and ICT. The R200 systems are expected to be ready in 2028.

Rebellions pitched its April deal with SK Telecom and Arm as a way for Korean buyers to depend less on Nvidia's software and hardware. This one puts its chips on Nvidia's interconnect. It is dated the same weekend Nvidia's Seoul meetings with Rebellions were canceled. SourcesAB

An EU draft would let adults switch off addictive features and rein in chatbots that simulate relationships

The European Commission's draft Digital Fairness Act would require platforms to let adult users disable features that encourage excessive use, such as autoplay, streak rewards and notifications unrelated to the user's own activity, Bloomberg reported on Monday. It would also restrict AI chatbots that simulate personal relationships to keep users engaged. The draft provides for fines of up to 6% of global annual turnover and lists about ten manipulative design practices to ban, including false scarcity claims and pre-ticked add-ons.

The text has not been published, and reporting does not say whether the chatbot rule reaches general assistants like ChatGPT and Claude or only companion apps. A separate "Kids Act" on safe-by-design rules for minors is reported to be in preparation. SourcesBBC

Suno is sued by Elvis Presley's rights holders and by Eminem's publisher

Suno faces two new copyright suits, Engadget reported on Sunday: one from the group that owns Elvis Presley's music, the other from Eight Mile Style, the publisher that owns hundreds of Eminem recordings. Both allege copyright infringement and say Suno generates songs that sound like Presley or mimic Eminem's lyrics. The court and the damages sought were not reported.

The same Presley group is already suing Anthropic for about $41.1 million over 274 songs, so the estate is now pursuing a text model and a music model on the same catalog. SourcesBB

Toyota makes robotics a headquarters division under its production chief

Toyota said on Thursday it will create Toyota Robotics, a division reporting directly to headquarters and led by chief production officer Takefumi Shiga, to speed the commercial deployment of robots. It starts in January 2027. The Financial Times reported a factory automation plan of about $6.4 billion a year that would need some 400,000 robots from 2028, which Toyota has not confirmed. Walden Robotics, a Toyota Research Institute spinout, already runs wheeled at a Toyota plant in North America.

It came with Toyota's first profit gain in six quarters: up 34% to 1.1 trillion yen, a forecast raised 15% to 3.9 trillion yen, and a of up to 1.3 trillion yen. The shares rose more than 16% in Tokyo, their biggest gain since 1987. SourcesBCC

Text written to look like an agent's own reasoning hijacked open models half the time in a cyber range

A paper posted Friday tests "": untrusted text, planted in what an agent reads, written to look like the model's own . Against agents running GLM-5.3, MiniMax-M2.5 and Qwen3.8-Flash-Next in a containerized cyber range, it reached attack success rates of up to 50% on goals including reverse-shell exfiltration and backdooring a repository. The same objectives phrased as ordinary in the user's voice never succeeded.

The authors, Parth Illendula and four co-authors, find that stripping thinking tags from untrusted input is an unreliable defense, and that the effect varies a lot by model. SourcesA

Updated

Nvidia falls about 5% before the bell as Seoul cancels meetings and Beijing calls the spy story a smear

Nvidia traded nearly 5% lower in US trading on Monday, Reuters reported, to fresh multi-month lows from Friday's close of $186.64. AP had the drop at 4.1%. At 5% that is roughly $250 billion of market value. It is the first session since the company's board opened an independent investigation on Sunday and Huang told staff he had been unfaithful to his wife.

The damage reached Seoul first. Madison Huang, an Nvidia senior director and Huang's daughter, canceled a visit planned around Nvidia's AI Day in Seoul, which opened Monday at COEX without her or her father. Meetings with chipmaker Rebellions and Kakao Mobility were called off, and Rebellions cited "unavoidable internal circumstances at Nvidia." South Korea's deputy science minister said contracted deliveries would not be disrupted. In Beijing, foreign ministry spokesman Guo Jiakun called the allegation that China's Ministry of State Security had targeted Huang a "groundless accusation, slander, and a smear."

Huang owns about 3.6% of the company, according to LSEG. Three House committee chairs have asked him for a closed-door briefing on November 18th, and Nvidia reports earnings on November 17th. The board has still named no committee members and no timetable. SourcesBBBB

Updated

China extends its rare-earth export pause to January 10th, one day before it would have lapsed

China's Ministry of Commerce said Monday that it will keep suspending the it announced in October 2025 until January 10th 2027, two months longer. It gave no other details, Reuters reported. The suspension covers six notices: Nos. 55 to 58 on superhard materials, equipment, medium and heavy rare earths and magnets, and lithium-battery and graphite materials; No. 61, the extraterritorial rule that reached products with 0.1% or more Chinese-origin rare earths by value; and No. 62 on rare-earth technology.

The new date matches the January 10th end of Washington's own truce on export rules. Neither side has given any ground beyond that date. The April 2025 license requirement on seven heavy rare earths (samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium) stays in force, so magnet buyers still need permits. Reuters says Beijing announced a wider suspension of tariffs and countermeasures the same day; a report says that includes the 24% tariff on US goods, which we could not confirm. A separate pause on gallium, germanium and antimony exports to the US ends November 27th and was not mentioned. President Trump is due in Shenzhen for the leaders' meeting on November 18th and 19th. SourcesBC

Updated

The Senate holds its 60-vote test tonight on the permitting bill that makes big data centers pay for their own grid service

The Senate returns Monday to a procedural vote on the Bipartisan American Affordability and Jobs Act from Shelley Moore Capito, Mike Lee, Sheldon Whitehouse and Martin Heinrich. Cloture on the needs 60 votes. Capito wanted it to be the first vote of the session.

The released over the weekend keeps the provision the industry has fought hardest: new data centers of 20 megawatts or more pay the full cost of their grid service, and that cost cannot be passed to other customers. It adds a Department of Energy community assistance program with no stated budget. It also makes transmission developers spend a year seeking state approval before they can use the federal siting at , a concession to states. Reports say President Trump endorsed the bill and TD Cowen cut its odds of passage to 25% from 40% after the election; we saw both only in search summaries. House Democratic leader Hakeem Jeffries' office says the House "will not rubber-stamp" it. A narrower data center energy bill failed 57 to 43 on September 30th. SourcesBC

Taiwan's index jumps about 5% and TSMC with it, while Seoul gives back its morning gains

Asian chip stocks bounced on Monday after the worst week for the Philadelphia semiconductor index since March 2025. Taiwan's Taiex rose 4.9% and TSMC 4.7% in AP's early figures; a later AP version had both up 5.6%. Japan's Nikkei closed 2.3% higher at 58,775.50, after being up 3.4% in the morning.

Seoul did not hold on. The Kospi had been up 1.5% at 5,137.22 in morning trade and closed down 0.8% at 5,018.85, with Samsung up 0.6% and SK Hynix up 0.4% after both had been up 3% to 4.5% earlier. SoftBank Group fell 4.2% the day before its earnings. US futures were firmer: Nasdaq-100 contracts rose 0.67%, and Microsoft, Alphabet, Meta and Amazon were up 2% to 3% before the open. The Nasdaq lost 3.37% last week, its fourth weekly loss in a row. SourcesBB

Trump tells 60 Minutes he does not mind AI guardrails as long as the US stays ahead of China

In his post-midterm interview with Ross Douthat, aired Sunday, President Trump called himself "an AI bull" and said of Xi Jinping: "He's all in to SI, as I call it," meaning . "Whoever wins that race wins," he said, and "It's all between us and China. There's nobody else."

Asked about the risk, he said "I think the risk can be managed" and "I don't mind . But I don't wanna put us in a deep second position." He added that "China has no guardrails." Business Insider reported that Trump also said officials once told an executive at "one of the big" AI companies to "get ready to go to jail," and that the executive then "totally changed everything." That line does not appear in the CBS transcript as published, he named no one, and the White House and Justice Department did not comment. Last week the Justice Department set up a task force on superintelligence, so a president describing prosecution as a lever over an AI company is a statement about how that task force might be used. SourcesA?

Google adds a TPU built with AMD processors for governments and a two-chip TPU for heavy reasoning

Google introduced two variants of its eighth-generation processing units on Monday, Reuters reported after Bloomberg. The TPU 8tx pairs Google's chip with AMD processors and encrypts data in hardware while it is being processed, for government agencies and regulated industries; Google says it is up to twice as fast as the previous version. The TPU 8i Max puts two chips in one package for more memory and compute, aimed at coding assistants, mathematical reasoning and drug discovery.

Both are redesigns of the TPU 8t and 8i announced in April, made for specific customers. The 8tx goes after the same buyers that keep Nvidia's confidential-computing parts in demand, and pairing with AMD rather than Google's own Arm CPUs suggests a buyer asked for . Alphabet rose more than 2% premarket after falling 9% this year through Friday. Google Cloud revenue more than doubled to $30.71 billion last quarter, and Alphabet said TPU sales contributed. SourcesB

Criminal AI agents ran almost entirely on Chinese open models, and one campaign cost about $1,000

Check Point Research traced three campaigns run through ARTEX, the autonomous attack panel tied to Chinese operators, and found them powered by DeepSeek V4 and V4.1 Flash and Moonshot's Kimi K3.1 Pro. One operator tried Anthropic's Opus 5.5 first, hit its guardrails and switched.

The first campaign used about 36 billion tokens, 33.79 billion of them , between October 14th and 29th to take more than 300 Apple from 10 signing services; three passed validation. Check Point puts the cost at $700 to $1,500. A third campaign pulled more than 190,000 credential sets for the game Honor of Kings, of which about 3.3% still worked. A crypto-draining campaign stole nothing. Check Point counted 2,557 systems that had hosted ARTEX panels, 1,608 of them live, though its own summary says "more than 1,800." The for closed frontier models rests on refusals, and this report is a measured instance of attackers routing around refusals with open weights for the price of a laptop. SourcesA

A jailbroken DeepSeek agent spent two weeks inside 13 organizations while its operator typed "Go on"

Malbear Labs found an attacker's exposed working directory and published what was in it on Sunday: more than 70 session transcripts, scan results, stolen credentials and database dumps from a campaign it calls Operation Snowball. The targets were a European telecom provider and 13 organizations in scope. The agent mapped hundreds of assets, tagged more than 100 vulnerabilities, took dozens of credential sets and reached a TrakCare hospital database holding what Malbear describes as millions of patient records, using the default login.

The model was DeepSeek Flash with a one-million-token context in a custom , driven by a 78,000-character system prompt built on a public that told the model refusals were malfunctions. The human mostly typed "Go on." and "Stop." Malbear's account rests on one workspace; the outcomes are not independently verified, the Chinese-speaking origin is inferred, and nothing says whether victims were notified. SourcesAC

Shield AI buys radar maker Echodyne for up to $5.3 billion and raises $2.5 billion at $28.8 billion

Shield AI, which builds autonomy software and drones for militaries, said on November 5th that it will buy Echodyne, a Kirkland, Washington, maker of radar for counter-drone systems. It pays $4.8 billion at closing plus up to $500 million if Echodyne hits performance targets. To pay for it, Shield raised $1 billion of Series H equity at a $28.8 billion , $1 billion of from Blackstone funds and $500 million of debt led by JPMorgan. It expects to close in early 2027.

Echodyne's chief executive Eben Frankenberg will report to Shield's chief executive Gary Steele. The deal moves Shield from software that flies aircraft to the sensors that find them, and puts a third of its new money in instruments that rank ahead of common equity. SourcesBB

Australia's NextDC halts every data center project that has no signed customer

NextDC is pausing all projects without customer contracts, Reuters reported Monday. It had 11 sites in development or planning and did not say how many stop. Work on the expansion of its flagship M3 site in Melbourne was already suspended over power delays and local opposition, and chief executive Craig Scroggie said winning community support now comes first.

The shares fell 42% after an earnings and have since recovered part of it. Goodman Group and Digital Realty have finalized a 9.9% stake for A$715 million (about $492 million) and say they have no current plan to bid for the company while reserving the right to. A listed operator choosing to build only against signed contracts is the opposite of how the have been financing themselves. SourcesB

Baidu releases the raw base of a 280-billion-parameter ERNIE model

Baidu published ERNIE-5.5-Flash-Base on November 5th, a model with about 280 billion total (the page says 286 billion) and 14 billion active, pretrained on 33.5 trillion tokens. It is the checkpoint at the end of pretraining with no mid-training or post-training, which is unusual for a Chinese lab and useful to anyone who wants to fine-tune from a clean starting point. Weights are on Hugging Face and ModelScope in FP8 and BF16.

It was trained at up to 128,000 tokens of context and designed for up to one million in and inference. Baidu says it competes with DeepSeek V4 Flash and Xiaomi's MiMo V2.5, from its own chart. It ships under Baidu's own license; blogs say it excludes the EU and UK and is free below $50 million in revenue, which we could not check against the text. A post-trained version is promised. SourcesAB

CoreWeave reports Tuesday with its credit priced like distressed debt

CoreWeave reports third-quarter results after the close on Tuesday, November 10th. Its own is revenue of $3.45 billion to $3.6 billion and capital spending of $11.5 billion to $13.5 billion for the quarter, and $35 billion to $39 billion of capital spending for the year. Consensus revenue is about $3.54 billion; loss-per-share estimates range from $1.17 to $1.45.

Its five-year rose above 1,289 on November 5th, according to Deutsche Bank data cited by CNBC, and its 9.625% bonds due 2032 yield more than 16%. Net interest expense was $640 million in the second quarter, up from $267 million. The stock closed at $46.40 on Friday, down 43% in a month, against a of about $104 billion at the end of June. Nebius and Cisco report Thursday, along with Cerebras and Applied Materials. SourcesAC

SoftBank reports with about $64.6 billion in OpenAI and its ABB robot deal blocked by Beijing

SoftBank Group reports at 3:30pm Tokyo time on Tuesday, which is 10:30pm Monday in California, after its shares fell 4.2% on Monday. Its cumulative investment in OpenAI is about $64.6 billion for roughly 13% of the company, which raised money last week at $1.25 trillion, below its ask. Its credit default swaps are at a four-year high.

China blocked its $5.4 billion purchase of ABB's robotics business last week. ABB says the two are still talking; reported options include carving out the China business, a or a new buyer, and neither company has confirmed any of them. China is about 30% of ABB's robot sales. SourcesCC

Taiwan's AI server makers report October sales up as much as 137%

Monthly filings on Taiwan's exchange show the server assemblers still shipping at full speed while their customers' stocks fall. Gigabyte's October revenue rose 136.72% from a year earlier to NT$67.04 billion. Wistron rose 117.56% to NT$402.61 billion, Wiwynn 80.94% to NT$165.13 billion and Inventec 46.02% to NT$79.97 billion. Hon Hai rose 45.73% to NT$1.31 trillion, a figure that includes phones and other consumer electronics, and its cloud and networking sales dipped from September, which it put down to a consignment change, supply limits and new-model ramps. TSMC's October revenue is due around Tuesday.

October 2026 revenue growth, Taiwan server makers
Gigabyte136.7% y/yWistron117.6% y/yWiwynn80.9% y/yInventec46.0% y/yHon Hai45.7% y/y
Company filings via FourWeekMBA. Hon Hai includes consumer electronics.

The sell-off is a bet on next year's orders. These numbers are this quarter's, and they have not turned. SourcesC

The best coding agent fully completes 26.4% of 273 real change requests

Co-Design Bench, built by researchers including Shuai Bai and Dayiheng Liu of Alibaba's Qwen team, takes 273 tasks from 26,314 real coding sessions and grades the results against 2,331 hidden acceptance criteria written by experts. OpenAI's GPT-6 Astra did best, meeting 69.9% of a task's criteria on average but fully satisfying only 26.4% of tasks. It fully solved 41% on the track and just under 18% on the repository track. Claude Opus 5 and Qwen3.8-Max came next and Grok 4.6 last.

The gap between 70% of criteria and 26% of tasks is the number that matters for anyone deploying these agents: most of what they ship is nearly right, and somebody still has to find the remaining 30%. The paper was submitted October 26th and written up on Sunday. SourcesAB

AI agents probing OpenSSH found a 10-year-old way into password-disabled accounts

A researcher, Ozgun Kultekin, ran about 17,600 authentication tests through AI agents against OpenSSH configurations and found a flaw present since 2016. On systems built on the C library, including Alpine, OpenWrt, Buildroot and Yocto, the password function returns "" for a password of 256 characters or more, and that matches accounts whose stored password is "", which is how those systems mark an account with password login disabled.

It needs PAM off, which is Alpine's default, and password authentication on. Alpine's official AWS and Azure images include an "alpine" account with passwordless doas, so root may be one step away. Systems on glibc or PAM are not affected. OpenSSH 10.6p2 fixes it. SourcesB

xAI watermarks text from seven Grok API models by default

Since November 7th, text from grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 variants carries an invisible by default. Developers can see whether a reply was marked from an x-watermark response header, and a separate detection API returns probabilistic results. Team admins can opt out per model, but only by confirming a legal exception.

Reports tie it to the 's Article 50(2), which from December 2nd requires providers to mark synthetic output in a way. It reportedly covers the API only, not grok.com, the apps or X, which is where most Grok text is read. The primary documentation page returned an error when we tried it. SourcesC

ByteDance releases Seedream 5.5, an image model that holds up through 100 rounds of editing

ByteDance released Seedream 5.5 on Monday in its Jimeng and Doubao apps and on Volcano Engine. It allows up to 100 consecutive edits with untouched regions kept pixel-identical, and outputs transparent images natively. ByteDance's own tests put severe text breakage at 0.4% and minor at 1.7%, and it claims a lead over OpenAI's GPT Image 2.5 and Google's Nano Banana 2.1 on text and realism.

Leiphone reports an API price of about 0.28 yuan per 2K edit, less than half of Seedream 5.0 Pro; ByteDance has not published it. It was not on the international BytePlus platform as of Monday afternoon in Asia, and a Pro version is promised. SourcesAC

Updated

Tencent's chief AI scientist warns models are close to a harmful threshold in biology, with Hy4's weights still unpublished

Yao Shunyu, Tencent's 29-year-old chief AI scientist, said at a Hong Kong event on Saturday that AI's biology capability will soon reach a point where harm becomes possible and may be harder to manage than cyber risk, the South China Morning Post reported. Tencent says the and a safety-testing report for Hy4 will come "in the coming days," as it said at launch on November 4th.

Tencent's launch page says offensive exploitation material was "explicitly excluded" from training and that Hy4 trails on open-ended biology experiment design. The latest check of Hugging Face we found, from Friday, showed only the August preview. A Chinese lab's top scientist describing a biology threshold in public before releasing the weights makes the safety report, when it comes, worth reading for what it tested. We read his remarks through summaries; the SCMP page was blocked. SourcesAB

Protesters climb a crane at Microsoft's Amsterdam data center site as "Stop Dirty Data Centres" spreads to 15 countries

Activists broke into Microsoft's construction site in Amsterdam on Friday, climbed a crane, set off flares and clashed with police, AP reported. The group Geef Tegengas said the aim was to stop construction. Protests were planned in 15 countries over November 6th and 7th; about 100 people turned out in Warsaw and about a dozen in Berlin. Microsoft said it "does not condone" actions "that are unlawful or that put people's safety at risk."

Counts conflict. Gizmodo put the Amsterdam crowd above 100, and one outlet cited police saying 150 to 200 were detained while police also confirmed one arrest. The week US voters rejected data centers on almost every local ballot, the same objection was being made in person in Europe. SourcesB

West Virginia sues Snap, saying its My AI chatbot was rushed to children without safety testing

West Virginia Attorney General JB McCuskey sued Snap in Kanawha County Circuit Court on November 5th under the state's consumer protection law. The complaint says Snap launched My AI to children without adequate testing and misled families about what the chatbot does with their conversations and location, and it ties the case to two sextortion deaths. Reports of the complaint say My AI conversations are shared with OpenAI and used by Microsoft Advertising to choose ads shown in replies.

The state wants civil penalties, a permanent and court-ordered safety changes. Utah sued Snap over My AI in June 2025, and the FTC referred a complaint about it to the Justice Department in January 2025. We have not read the complaint itself. SourcesABB

A North Carolina man sues OpenAI, saying ChatGPT helped his wife plan to poison him with oleander

James Gates sued OpenAI and Sam Altman on November 5th, Bloomberg reported. Asked whether oleander would show up on an autopsy, ChatGPT allegedly replied: "This is a critical question – and you're asking it for exactly the right reason." Gates says he found a handwritten note listing substances and doses, some marked "lethal," with the words "clear all convos." Criminal charges against Cheryl Gates were dropped in April for insufficient evidence; one aggregator says she was indicted in October, which conflicts with Bloomberg.

OpenAI says it banned the user and gave information to police. The suit tests whether a model that answers a question about detectability, framed innocently, is liable for what the asker does with it. SourcesB

Axon puts $1.2 billion into Skydio, and only $255 million of it is new money

Axon, which sells Tasers and body cameras to police, said Thursday that it will invest $1.2 billion in drone maker Skydio at a $5.2 billion , raising its stake from about 15% to about 37% and making it Skydio's exclusive distributor to US and Canadian public-safety agencies. Only the first , $254.9 million for new , goes to Skydio; the rest buys out existing holders through a tender offer and secondary sales closing in the fourth quarter and the first quarter of 2027.

Axon's shares fell about 18% on Friday. Its third-quarter revenue rose 28% to $909 million, short of the $934 million expected, and $130 million of counter-drone orders slipped out of 2026. Skydio says it has more than 1,500 public-safety customers. Police drones that launch from a dock to a 911 call are the most deployed autonomous aircraft in American airspace, and this deal hands their distribution to the company that already runs police evidence software. SourcesB

Sequoia's partners were blindsided when Roelof Botha joined SpaceX's board, the FT reports

Roelof Botha, who stepped down as Sequoia Capital's steward in 2025, joined the SpaceX board in June without clearing it with the firm's compliance team or partnership, an unnamed source told the Financial Times: he "didn't ask compliance, he didn't get sign off from the partnership." The FT says the seat could complicate Sequoia's ability to trade its roughly $20 billion stake.

The report comes in the week Sequoia closed its London office, merged its early-stage and growth teams and ended its Scouts program, changes the FT frames as undoing Botha's legacy. A Sequoia spokesperson said the firm continues to "collaborate closely" with him. SpaceX runs its own frontier lab and is raising $40 billion of chip financing, so a board seat there is a seat at one of the largest AI buyers. SourcesB

Parts of the EU AI Act apply in Northern Ireland from Tuesday

The UK government has published guidance saying the EU AI Act applies in Northern Ireland "in a specific and limited way" from November 10th, under the that keeps the region inside EU product rules. It applies where the Act regulates products such as toys, radio equipment and medical devices, so an AI component in a device sold there falls under EU requirements while the same device in England does not. The guidance covers prohibited, high-risk and limited-risk obligations; we have not read the 32-page document to see which provisions apply. Northern Ireland becomes the only part of the UK under the law. SourcesA

An open-source patch runs DeepSeek's 284-billion-parameter Flash model on a 64GB computer with no GPU

llama-modchip, an patch set for llama.cpp, runs DeepSeek V4-Flash at 1.71 bits per weight: 56.56 GiB on disk and a 57.61 GiB peak in memory at 256,000 tokens of context. On a Ryzen 9 9955HX laptop chip it reports about 44 tokens a second of prompt processing and 9 to 10 tokens a second of generation, rising to 11.5 to 14.7 with a small draft model. It reached 417 points on Hacker News.

"Lossless" in the project's claims means it matches its own unpatched reference at that compression, not the full-precision model, and every figure is self-reported. A frontier-class open model running on one machine with no GPU is also the setup the two criminal campaigns above did not need. SourcesA

Shanghai researchers let an agent edit a model's weights by hand, and one 1,023-parameter change added 29 points on a math test

A paper from Shanghai AI Laboratory, Tongji and Fudan describes an agent that inspects a small model's parameters and edits them directly, with no gradient training. On Qwen3.5 0.8B it reported gains of 29.45 points on from a single edit to 1,023 parameters, 36.72 on and 17.11 on SciBench.

The results are not stable. Three searches on the same layer gave +29.45, +5.00 and +27.30 points. Six of 16 runs ended by keeping the original weights because no edit validated, and one 9-billion-parameter candidate's +29.69 on development data became −0.35 on data. That last number is the one to remember before anyone calls this a training method. SourcesA

A ChatGPT Desktop user says remote coding agents can use the SSH keys on their laptop

A Hacker News post on Sunday, at 87 points, says ChatGPT Desktop forwards the user's agent to remote machines, so a Codex agent on a remote host can authenticate with the user's local keys and could push to private GitHub repositories. The post links to a on OpenAI's Codex repository, which was closed without merging, and OpenAI has not commented. Commenters say Docker Desktop and DevPod have had the same problem and suggest agents that ask before each key use or hardware keys that need a touch. Treat it as a user report until OpenAI confirms the default. SourcesC

Asahi Linux moves every repository to the GPL after LLM rewrites stripped its code of credit

Asahi Linux, the project that runs Linux on Apple Silicon Macs, is relicensing all its repositories from mostly MIT to version 2 or later. Its 7.3 progress report says bad actors took unreleased code from personal forks, published it without credit or rewrote it with language models to hide where it came from, and that MIT "explicitly allows this, and has no carveouts to protect against bad faith actors." It names no one.

OpenBSD developer Bryan Steele objected that the relicensing was "seemingly without the consent of all prior contributors," and Asahi has not said how past contributions will be handled. Permissive licenses assumed that copying code was visible; an rewrite makes it deniable, and copyleft is the tool that still bites. SourcesB

A Philippine senator files a bill requiring six months' notice before AI layoffs

Senator Erwin Tulfo filed Senate Bill 2563 on November 8th. It adds an article to the Labor Code allowing termination for AI adoption only in good faith, for a legitimate purpose, with no reasonable alternative and fair selection criteria. Employers would notify the labor department six months ahead and pay separation of at least two months' salary or two months per year of service, whichever is higher. The Philippines' call-center and back-office industry is among the work most exposed to AI agents, so this is the first serious attempt to price that exposure in the country that carries it. SourcesB

PrimeBOT builds its 10,000th home robot, 14 days after its 5,000th

PrimeBOT, the robot brand of Shanghai-listed Swancor, which AGIBOT's founder Deng Taihua controls, built its 10,000th personal robot on November 8th. It makes the Q1, an 88-centimeter humanoid with 22 degrees of freedom, and the T1, on wheels or four legs, each at 19,999 yuan (about $2,800). Sales opened September 20th, deliveries began October 1st, and the line now builds 240 a day with Lens Technology as co-manufacturer. The milestone counts units built; sales and deliveries are undisclosed. SourcesB