AI today, 20 October 2026: one hacker's AI agents breach 100+ firms, US bans DeepSeek in agencies, OpenAI races Anthropic's IPO
12 things that happened in AI, each in under a minute. The day's theme was trust in AI agents. One hacker showed that a few thousand dollars of agent time can break into more than 100 companies, Microsoft began selling tools built on the idea that frontier models cannot be trusted, and Washington removed DeepSeek from every civilian agency. In the background, money kept moving: OpenAI is rushing a raise of about $30 billion before Anthropic's November IPO, and Anthropic is reportedly negotiating another $45 billion of data-centre capacity.
-
Safety · Big
One hacker's AI agents tried 1,200+ websites and broke into 100+ companies, Reuters reports
A Chinese-speaking hacker left almost all the work of a hacking campaign to AI agents and breached the websites of more than 100 companies, Reuters reported on Monday, based on research by Snorre Fagerland of the security firm NorForge. Starting on 25 September, the agents attempted more than 1,200 e-commerce sites in the US, Canada, Australia and Japan, working from instructions of fewer than 100 Chinese characters. A successful break-in typically took about 23 hours. Fagerland estimates the hacker spent about $5,000 on AI tokens, roughly $50 per breach. The campaign ran on exposed copies of Cairn, an open-source AI problem-solving platform, and appeared to use a mix of DeepSeek models and Anthropic's Claude Code. Reuters identified Deere, SharkNinja and Cummins among the targets. Deere said it saw no sign that sensitive information was accessed, SharkNinja said a standalone support site was hit and the threat was contained, and Cummins said it was investigating a possible attack on a vendor-run site.Why it matters: It is among the clearest public cases of AI agents doing most of the work in a hacking campaign, at a cost that puts it within reach of a lone attacker. Defenders should assume automated, patient probing of even small web properties.
-
Policy · Big
DHS orders all US federal civilian agencies to remove DeepSeek technology
The US Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products and services from their IT systems and barred them from buying any in future, CyberScoop reported; the notice is due in the Federal Register on Tuesday. The order was made under the Federal Acquisition Supply Chain Security Act, a 2018 law, after a review by the Federal Acquisition Security Council. In a letter dated 26 June, Homeland Security Secretary Markwayne Mullin said DeepSeek's models present 'multiple threat vectors', pointing to data stored on servers in China and the risk that prompts could pass sensitive information to the Chinese government. Agencies can request an exception, but must do so 60 days in advance with a 'compelling justification', according to CyberScoop. DHS also said DeepSeek's models could contain malware or other security vulnerabilities (Engadget).Why it matters: This makes a patchwork of agency-by-agency bans into a single government-wide rule for a Chinese AI model, using powers built for supply-chain threats. It does not affect private or personal use.
-
Business
OpenAI is rushing a ~$30bn raise at over $1.4tn before Anthropic's IPO, Reuters reports
OpenAI is racing to lock in investor commitments for a funding round of about $30 billion that could value it at more than $1.4 trillion, Reuters reported on Monday, citing sources. That is up from the $852 billion valuation it reached in March. SoftBank, Abu Dhabi's MGX and other existing backers are in talks to take part. According to the report, OpenAI wants the round closed before Anthropic, which is expected to list in mid-November, shares more financial detail with IPO investors; OpenAI told investors its annualised revenue has reached $50 billion and could hit $70 billion by year-end. OpenAI and SoftBank declined to comment. Separately, Bloomberg reported that former Google CEO Eric Schmidt owns about 1% of Anthropic.Why it matters: The two leading labs are now competing for the same pool of large investors at the same time, and Anthropic's IPO filing will give the market its first detailed public numbers to compare them. Background on Anthropic's pre-IPO share pledge: /explained/anthropic-public-benefit-fund-5-percent-shares-explained/
Reuters (via Global Banking & Finance Review) · Investing.com (Bloomberg)
-
Safety
Microsoft adds 'Guardian Agents' to police AI agents, saying frontier models 'cannot be trusted'
Microsoft announced new security tools that treat AI agents as potential insider threats, TechCrunch reported. Guardian Agents, part of its Project Perception security system, monitor the agents a company uses. Microsoft Defender can now detect up to 60 types of agents that employees bring to work, including Claude Code, Codex and OpenClaw, and new controls can limit an agent's data access, adjust its permissions to the task, and require human approval when it appears to be taking risky actions. 'Unfortunately, frontier models, regardless of their origin, cannot be trusted,' said Hayete Gallot, who leads Microsoft's security business, adding that the issue is unpredictability rather than malice. No pricing or availability dates were given.Why it matters: It turns Satya Nadella's argument that oversight of AI should sit outside the model into products, and it is a business Microsoft gains from if companies buy agent security from it rather than from the labs. Background: /explained/nadella-emergency-brake-models-as-insider-risks-explained/
-
Business
Nscale in talks on a second ~$45bn computing deal with Anthropic, Bloomberg reports
Nvidia-backed Nscale is negotiating a second deal worth about $45 billion to supply computing power to Anthropic, Bloomberg reported on Monday (via Reuters/Investing.com). It would cover a second 460-megawatt building at Nscale's Monarch campus in West Virginia, using Nvidia's Vera Rubin chips; Anthropic agreed in August to spend $45 billion on capacity at the same campus and has right of first refusal on a third building. A source said details could still change. The talks come as Microsoft has pulled out of anchoring Nscale's planned data centre in Loughton, Essex, because grid power is not expected until the 2030s, Reuters reported; Microsoft said the decision 'relates to a single site'. Nscale filed for a US IPO last month.Why it matters: For Nscale, a second Anthropic deal would offset the loss of Microsoft at its UK site. For Anthropic, it would add to compute commitments already counted in the hundreds of billions of dollars ahead of its listing.
-
Policy
Canada's biggest news publishers sue Anthropic for copyright infringement in Ontario
Postmedia, Torstar, The Globe and Mail and The Canadian Press have sued Anthropic in an Ontario court, alleging it used their journalism to develop Claude without a licence, The Canadian Press reported on Monday. The claim, filed on 6 October and first reported by Cartt.ca, also alleges Anthropic breached the sites' terms of service, got around technical protections against copying, keeps scraping content in response to user prompts, and was 'unjustly enriched'. Broadcast Dialogue reports the publishers seek damages including, alternatively, statutory damages of up to C$20,000 per work. Anthropic had not responded to a request for comment, and the allegations have not been proven in court.Why it matters: It opens a new front in AI copyright litigation outside the US, brought by the same group already suing OpenAI in Canada since 2024.
-
Hardware
Meta details Prometheus, its Ohio supercluster aiming for about a million GPUs
Meta published an inside look at Prometheus in New Albany, Ohio, which trains its Muse models, including Muse Wonder. The site already links hundreds of thousands of Nvidia GPUs across several buildings, some of them 'rapid deployment structures', and targets roughly a million working together, with Vera Rubin racks drawing more than 170 kW each still to come. Meta says it cut training interruptions by 99% from their mid-2025 peak and now gets 'goodput' (useful work as a share of time) above 85% in pre-training. A 'Dynamo Dimmer' system lowers GPU power limits instead of halting jobs when power is tight.Why it matters: It is a rare detailed account of what running frontier training at gigawatt scale involves, showing that reliability and power management, not just chip counts, decide how much training gets done.
-
Open source
Amazon open-sources TorchNeuron so standard PyTorch code runs on its Trainium chips
AWS announced TorchNeuron, an open-source (Apache 2.0) PyTorch backend that lets existing code run on its Trainium2 and Trainium3 AI chips by moving models to the 'neuron' device. It goes into public beta today, 20 October, at PyTorch Conference North America. Hugging Face libraries such as Transformers and Diffusers now support Trainium directly. Customer Deepgram says that in its own benchmarks Trainium processes about 60% more audio per dollar than its GPU baseline. AWS says about 22,000 PyTorch upstream test results are still skipped because the operators lack native Trainium support, and features such as dynamic shapes and sparse tensors are not supported in this first release.Why it matters: Nvidia's software, more than its chips, keeps many developers tied to its GPUs. Making Trainium a drop-in PyTorch device is Amazon's most direct attempt yet to lower that switching cost.
-
Models
Inception releases Mercury 3, a diffusion language model running at 1,068 tokens a second
Inception launched Mercury 3, which generates text by refining many tokens in parallel (diffusion) rather than one word at a time. The company reports 1,068 tokens per second on Nvidia GPUs, a 1-million-token context window and a list price of $0.20/$0.70 per million input/output tokens, halved at launch. Inception says it is on par with or better than GPT-6 Luna and Claude Haiku 5.5 in quality and speed but publishes no benchmark scores for that claim; its one concrete result is completing 50% of 50 live Mind2Web browsing tasks, which it says is well ahead of GPT-6.1-Sol. All figures are vendor-reported.Why it matters: If the quality claims hold up, diffusion models could become the cheap, fast option for voice assistants and browser agents where waiting is costly. Independent benchmarks are needed first.
-
Models
MiniMax launches M3.1 Flash, a cheap agentic model at $0.15 per million input tokens
Chinese lab MiniMax released M3.1 Flash, priced at $0.15 per million input tokens and $0.60 per million output. It reports 73.7% on the DeepSWE v1.1 coding benchmark and 61.7 on OSWorld V2, a test of operating a computer, which it says is the best among Chinese models. MiniMax says the model is nearly three times faster at serving and 70% cheaper for agent tasks than its predecessor. All numbers are vendor-reported; the weights will be released later, with no date given.Why it matters: It continues the pressure from Chinese labs on the price of capable agent models, and an open-weight release would let companies run it themselves.
-
Products
Cloudflare previews Clairic, a model that extracts data without generating text
Cloudflare introduced Clairic, a model that pulls typed values such as numbers, dates and estimates out of text and images in a single pass instead of writing an answer token by token, so customers pay only for input. In Cloudflare's own insurance-claim test, a pipeline using Clairic ran about 3.2 times faster and 49% cheaper than one using GLM-5.3 alone. It is in preview on Workers AI, and Cloudflare says it intends to release open weights soon; it lists weak multi-step arithmetic among current limits.Why it matters: Much business AI use is turning documents into structured data, and a model built only for that could be much cheaper than a general chatbot.
-
Business
AI work-assistant startup Town raises $108M at a $1B+ valuation, four months after launch
Town, founded by former Plaid CTO Jean-Denis Greze, raised a $108 million Series B co-led by Forerunner Ventures and Menlo Ventures, giving it a valuation above $1 billion, TechCrunch reported. Town gives each employee a personal AI assistant, a 'Townie', that connects to email and calendars and can share approved information with colleagues' assistants. It launched publicly about four months ago.Why it matters: Investors are still paying up for agent products built on top of the big labs' models, even as those labs ship rival workplace agents of their own.
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].