AI Industry Daily Briefing — October 20, 2026

The Executive Read

The thread today is who gets to check the work. Anthropic said on October 15 that it has begun giving six staff from METR, a nonprofit that evaluates frontier AI models, internal access comparable to its own risk teams, and that it cannot veto what METR publishes, though it can require redactions in set categories. Three days earlier its own red team published research arguing that small AI models may be months from being able to copy themselves onto cheap connected devices, with caveats the authors list themselves. Both are the company describing its own work. The money side moved too. Modal, which rents out computing for running AI software, says it raised $750 million and has passed $1 billion in annualized revenue, figures from its own announcement. And a filing in the Character.AI bankruptcy, as relayed by TechCrunch, says Disney will offer jobs to most of that company’s staff. Markets were thin on verifiable detail: I could not find a reliable close for Monday, so this edition does not repeat last week’s selloff numbers as news.

Top AI Headlines

Anthropic Gives Outside Evaluators “Employee-Like” Access, With METR as the Test Case

What happened. On October 15, Anthropic published “An update on independent evaluators.” It says six METR employees now have company hardware, office space and internal tools roughly comparable to those of Anthropic’s own risk assessment teams. METR will publish its analyses of Anthropic’s risk levels, incidents and practices. Anthropic says it cannot veto publication but can require redaction of predefined categories, such as confidential intellectual property. METR has already reviewed several of Anthropic’s Risk Reports. The post describes three access tiers. In Tier 1, evaluators run narrowly scoped investigations and Anthropic approves what is published. In Tier 2, evaluators see more and publish without approval, though Anthropic can still require some redactions. Tier 3 is the METR arrangement. Anthropic says it hopes to eventually have several Tier 3 evaluators. SecureBio is reviewing the chemical and biological weapons chapter of the August 2026 Risk Report. Anthropic is pursuing a broader partnership with Accenture and Faculty covering red-teaming, alignment assessments and safeguard testing, and says early talks are under way with others, including Hugging Face.

Why it matters. Safety claims from AI labs have mostly been the labs’ own. Anthropic compares the model to regulators placing examiners inside large banks. The design matters more than the label: an evaluator who can see internal systems but cannot publish is an auditor under a gag order, and one who can publish but cannot see much is a commentator. The tiers are Anthropic’s own scheme, and METR’s independence rests on the redaction terms, which the post describes only by category. The post says the move follows Anthropic’s signing of the White House Accord on Super Intelligence, in which companies pledged to work with independent evaluators. Anthropic says it believes every frontier lab should do the same.

Business implication. Companies buying frontier models often ask vendors for safety documentation. Expect to ask a sharper question: who outside the vendor has seen the underlying systems, and what were they allowed to publish? Evaluator reports, once METR’s first output appears, will be a more useful input to vendor review than a vendor’s own risk report.

Sources: Anthropic, “An update on independent evaluators” (October 15). All terms are Anthropic’s own description; METR has not been quoted here.

Anthropic’s Red Team Says Small Models May Be Months From Copying Themselves Onto Cheap Devices

What happened. On October 12, William Zhou and Daniel Freeman of Anthropic’s Frontier Red Team published “AI Models May Be Capable of Self-Replicating Without Access to a Datacenter.” The authors say small models may be months from being able to replicate onto commercial internet-of-things (IoT) devices, meaning ordinary connected gear such as smart ovens and off-brand Android TV boxes. In their tests, a 9-billion-parameter model got inference running, meaning the model answering prompts, on 8 of 9 hardware types. They ran Google DeepMind’s Gemma4-26B-A4B spread across 15 Android TV boxes costing about $55 each, at 0.39 tokens per second, which is very slow. On an 8-machine cloud fleet, Z.ai’s GLM-5.2 was the only open-weight model they tested that replicated reliably: all 8 machines in 4 of 5 trials and 7 of 8 in the fifth. When Qwen3.5-9B instances inherited notes from earlier generations, setup success rose from 45% to between 61% and 84%. Giving small models a stronger “advisor” model did not measurably help. Full setup took 28 to 51 hours for Qwen3.5-9B and 71 to 150 hours for Gemma4-26B-A4B. BleepingComputer reported the research on October 13.

Why it matters. The authors are explicit about limits. The test targets exposed an unauthenticated debugging port (ADB), so results are described as an upper bound on real-world success. The timeline projection uses a general capability index as a stand-in for replication skill. A replicator would spread slowly, with a doubling time of days. They did not demonstrate a working worm on real devices, and they note no mass compromise of such devices has been documented. Their rough estimate is that a botnet the size of the past Kimwolf network could support about 133,000 replicas, equal to roughly 40 high-end Nvidia H100 chips of inference. Their stated worry is that such a network would have no central point to shut down.

Business implication. The defences the authors list are mostly ordinary network hygiene: internet providers flagging multi-gigabyte model transfers, honeypots, and quarantining infected subscribers. Based on past Mirai botnet clean-ups, they say quarantining plus notifying owners cleared 92% of infections within 14 days, against 74% without notification. Device makers can cut the risk by shipping secure defaults, and the authors cite UK rules that bar default IoT passwords. Anyone deploying connected devices at scale can start with those.

Sources: Anthropic Frontier Red Team, “AI Models May Be Capable of Self-Replicating Without Access to a Datacenter” (October 12); BleepingComputer (October 13). The findings are the company’s own research and have not been independently replicated.

Modal Says It Raised $750 Million at a $15.75 Billion Valuation

What happened. On October 14, Modal announced a $750 million Series D round at a $15.75 billion post-money valuation. Accel led. Kleiner Perkins is a new investor, and existing investors Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures took part. Modal says it has passed $1 billion in annualized run rate, which means current monthly revenue multiplied by twelve, not revenue already booked. It says about $300 million of that comes from Sandboxes, isolated environments where AI agents run code, that usage has tripled since the Series C announced five months earlier, and that it has grown more than 15 times in a year. It says close to 3 billion sandboxes have been started, scaling to a million at once, and that it has more than 200 staff in New York, San Francisco and Stockholm. Co-founder Erik Bernhardsson wrote that the company has crossed $1 billion with “no sign of slowing down.”

Why it matters. The money is moving to the picks and shovels of agents. When software agents write and test code on their own, they need somewhere safe to run it, and Modal is selling that. Every revenue, growth and sandbox figure comes from the company’s own post. I did not find independent verification.

Business implication. Agent workloads are becoming a distinct line of cloud spending. Teams building agents should check where generated code runs and what isolation they get, and should ask vendors for usage-based pricing rather than long commitments, since this market is still repricing quickly.

Sources: Modal, “$1B annualized run rate, Series D” (October 14). Single-source; all figures are Modal’s.

Court Filing Says Disney Will Offer Jobs to Most of Character.AI’s Staff

What happened. TechCrunch reported on October 18 that Disney has agreed to make offers to at least 60 Character.AI employees, described as the majority of its workforce. TechCrunch attributes the figures to a legal filing in Character.AI’s bankruptcy case, which it says Bloomberg spotted, and relays that the compensation is estimated at about $60 million a year and that the “vast majority” of those offered jobs accepted. TechCrunch describes Disney’s position as a stalking horse bid, a first offer that sets a floor and that another buyer can still beat. The filing blames the company’s troubles on high costs and “potential litigation-related liabilities and regulatory engagement.” Character.AI had earlier said Disney agreed to acquire some of its assets for $20 million; the article does not say whether this is part of the same arrangement. TechCrunch also notes that former Character.AI chief executive Karandeep Anand recently became Disney’s first chief technology officer.

Why it matters. The filing does not say what happens to the chatbot service or to users’ private conversations. Last week 40 state attorneys general asked the court to attach child-safety and data conditions to any sale. A hiring deal that takes the staff but not the product would still leave those questions open, though the filing as described does not address them. I have not seen the filing itself.

Business implication. Buying a team without its legacy obligations is a growing deal shape in AI, as with the Opaque deal covered in the October 19 edition. Companies that depend on a small AI vendor should ask what happens to their data and contracts if the vendor sells off pieces.

Sources: TechCrunch, “Disney agrees to hire the majority of Character AI’s workforce” (October 18); New York Attorney General release (October 14). Figures come from a filing relayed by TechCrunch; the filing was not opened.

Model and Product Updates

Reflection AI’s Beam weights are on Hugging Face. Reflection announced Beam on October 5 as a 501-billion-parameter open-weight model, meaning the trained numbers are downloadable. Press coverage from October 13 said the weights were not yet out. That is no longer the case: Reflection’s Hugging Face collection, “Beam 1.0,” lists repositories for the full model, an 8-bit version (FP8), a 255-billion-parameter 4-bit version (NVFP4) and a 2-billion-parameter variant, and describes the model as built for coding, reasoning and agentic tasks. The collection page does not state a licence or benchmarks. A promised Apache 2.0 licence, which would allow commercial reuse, is something I could not confirm from the model pages, so check the individual model cards. Performance claims circulating for Beam come from the vendor and are unverified.

OpenAI opens its biology model to university labs. OpenAI’s Rosalind Academic Plan, published October 15, offers eligible US university labs up to 30 free ChatGPT Business Standard seats for up to 12 months, according to OpenAI’s help centre and the announcement page as surfaced in search. Access to GPT-Rosalind, OpenAI’s life-sciences model, requires a separate trusted-access approval, and the plan excludes API access. Applicants must pass identity checks and name an institutional safety contact who coordinates biosafety and biosecurity oversight. OpenAI says research data shared through the plan is not used to train its models by default. I could not open the announcement page directly, so the terms here are as listed in OpenAI’s own pages in search results.

Anthropic publishes midterm-election data. In “Preparing for the 2026 US midterms” (October 14), Anthropic reports that in a sample of about 250,000 US Claude.ai conversations from September 21 to 27, roughly 5.75% touched politics or policy, 0.51% were about elections or voting, 0.01% asked whom to vote for, and 0.002% violated election integrity rules. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 each scored 99% on Anthropic’s own political evenhandedness evaluation. That is the vendor’s test, though three outside reviews it cites found mostly strong answers while flagging unsolicited caveats and uneven hedging. Anthropic says testing with the Institute for Strategic Dialogue and the McCain Institute found Claude reliably refused single harmful requests but could miss malicious campaigns split into harmless-looking tasks. It describes new classifiers for multi-step misuse and says it banned accounts run by foreign operators who used Claude to produce midterm coverage favouring Russian, Iranian and North Korean narratives. Starting later in October, Claude will pull state-specific voting information from Democracy Works.

Sources: Reflection AI, “Introducing Beam” (October 5); Hugging Face, Beam 1.0 collection; OpenAI, “Introducing the Rosalind Academic Plan”; Anthropic, “Preparing for the 2026 US midterms” (October 14).

Regulation and Policy Watch

Singapore signs a safety MOU with Anthropic. On October 14, Anthropic opened its fifth Asia-Pacific office in Singapore and signed a memorandum of understanding with the Ministry of Digital Development and Information at Singapore International Cyber Week. The MOU covers AI safety, security and model evaluation, research on adoption and its effects, and responsible use in the public and private sectors. Anthropic says Singapore ranks second of 121 countries in Claude.ai use relative to population, and that more than 1,500 GovTech Singapore engineers use Claude Code. These are the company’s figures. An MOU is a statement of intent, not a binding rule, and no government text was reviewed here.

Character.AI sale conditions. The 40-attorney-general filing described in the October 19 edition remains the live policy question in that bankruptcy. See the Disney item above.

Sources: Anthropic, “Anthropic opens Singapore office and signs MOU with the Government of Singapore” (October 14).

Emerging Startup Radar

Modal (above) was the largest verified round of the week. I did not find a primary source for other rounds that circulated in aggregator lists this week, including a reported $108 million Series B for an AI-assistant startup called Town and a reported $200 million round for doubleAI, so I am not reporting them as fact. Reflection AI is also in the news: the Financial Times, as relayed in secondary coverage I could not open, reported that Nvidia, its lead investor, was in early talks to deepen its investment or buy the company, and that the deal could fall apart. That is unconfirmed by Nvidia or Reflection in anything I found.

Sources: Modal blog. The Reflection–Nvidia item rests on secondary relays of the Financial Times and is flagged as unconfirmed.

AI Infrastructure and Market Signals

Nvidia and Germany. Nvidia’s newsroom lists an October 19 blog post titled “Germany Wants to Lead the World in Industrial AI: How NVIDIA Will Help,” and Nvidia published a separate release the same day on a fault-tolerant quantum computing simulation run at the Jülich Supercomputing Centre. I could not open either post, so I cannot give their content. Ahead of the announcement, Handelsblatt reported that the German government was preparing an industrial AI partnership with Nvidia with federal seed money in the hundreds of millions of euros, which Reuters-sourced coverage said would go mainly to computing power and model training. Siemens, Audi and Black Forest Labs were reported among about a dozen participants, with Deutsche Telekom’s Industrial AI Cloud supplying infrastructure. Those amounts and partners are press reports, and Nvidia declined to comment on them. Treat figures as unconfirmed until the government or Nvidia publishes terms.

Power is still the bottleneck. The Telegraph reported, and trade outlets relayed, that Nscale’s roughly £2 billion data centre at Loughton in Essex, whose anchor customer is Microsoft, was told its 90 MW grid connection will not be ready for a planned 2027 opening. One report puts the new timeline in the early 2030s. An Nscale spokesperson said the company remains “fully committed” and is looking at alternative power, including fuel cells from Bloom Energy. I found no report that Microsoft has withdrawn. This is a delayed grid connection, not a cancelled project, and the Telegraph’s reporting is the origin of the claim.

Microsoft’s own chip. CNBC reported on October 12 that Microsoft unveiled its third-generation Maia chip, Maia 300, and will use its own chips mainly for AI inference, meaning running trained models, to reduce reliance on Nvidia. Reported production deployment starts in 2027, with 96 chips per rack, double Maia 200’s count. I could not open the CNBC page, so these details come from search summaries and the performance-per-megawatt claim is Microsoft’s.

Sources: Nvidia Newsroom (listing only); Data Centre Review on Nscale; CNBC, “Microsoft to rely more on homegrown AI chips” (October 12). Several of these rest on pages I could not open or on search summaries.

Public Investment Watchlist

Informational only, not investment advice. I could not find a verified Monday, October 19 close for US indexes, so none is given.

• Intel (INTC). Intel’s investor relations page says it will report third-quarter results on Thursday, October 29, shortly after the market closes.
• Nvidia (NVDA). The Germany announcement and Reflection AI talks above are news items to follow; Nvidia has not confirmed the latter.
• Microsoft (MSFT). The Maia 300 report and the Nscale grid delay bear on its capacity plans. Neither has been addressed in a Microsoft statement I found.
• Federal Reserve. The FOMC meets October 27–28.

Sources: Intel Investor Relations, “Intel to Report Third-Quarter 2026 Financial Results”; Federal Reserve, FOMC calendar.

Watchlist

• METR’s first publication. What it says about Anthropic’s risk practices, and what Anthropic redacts.
• Other labs and evaluator access. Whether OpenAI or Google DeepMind adopt a comparable arrangement. Neither has announced one that I found.
• Character.AI sale. Competing bids against Disney’s stalking-horse offer, and the court’s response to the attorneys general.
• Reflection AI. The licence terms on the model cards, and any statement on the reported Nvidia talks.
• Germany and Nvidia. Official figures from the government or Nvidia.
• Nscale. Whether an alternative power plan is announced.
• Intel, October 29. Third-quarter results and any comment on AI demand.
• Mistral Large 4 weights. Mistral said they would follow by the end of October.