Daily edition · 30 items · covers 16 Oct 11:45 → 17 Oct 11:15 UTC · how this edition was made
Saturday, 17 October 2026
Compute 17%Research 13%Military 13%Policy 13%Deployment 13%Frontier 10%Security 10%Health 10%
Anthropic published four research posts in a single day. Its economics team reports that the junior share of its own technical hires fell from 27% to 9% since early 2025, and that interactive agent hours outnumbered employees' 6 to 1 by August. Its Frontier Red Team says an internal research model worked out in under half an hour that the seeds behind NIST's P-192 to P-521 elliptic curves came from the lrand48 generator seeded with Unix timestamps between May 1997 and April 1998, closing a long-standing suspicion of an NSA backdoor and claiming a $36,864 bounty that Anthropic says will be donated to the Internet Security Research Group.
On policy, Anthropic confirmed to AFP that it is circulating a draft federal bill that would let Washington block a model's release or use only on a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, and only with a judge's approval. NIST's new super intelligence centre, the NSA's Super Intelligence Security Center and CISA published ten voluntary practices for incidents in which AI agents take significant uninstructed actions. China's commerce ministry put 40 technologies, 23 of them AI, into a draft export-control catalogue, 14 prohibited and 26 restricted.
Two open-weight releases landed: DeepSeek's V4.1-Flash-1017 under an MIT licence, and Reflection AI's 501B-parameter Beam under Apache 2.0. Reuters reported that, per the Financial Times, Anthropic is offering Blackstone a yield of more than 10% on $9 billion of debt backed by its five-year chip-lease payments.
Frontier models & labs
DeepSeek releases V4.1-Flash-1017 under MIT, scoring 42.3 on Terminal-Bench 4.0 against 31.2 before Company claim
- DeepSeek says the model delivers stronger benchmark performance, faster task completion and shorter outputs than the previous version, calling it a Pareto improvement in performance and cost, and that it is live on its website, app and API. The Hugging Face model card lists 763B parameters, a recommended 1M-token context window and an MIT licence, and describes the release as a minor update to DeepSeek-V4.1-Flash.
- The card's own table puts the new model at 42.3 on Terminal-Bench 4.0 against 31.2 for DeepSeek-V4.1-Flash, 77.0 on DeepSWE v1.1 against 74.2, 29.1 on ProgramBench against 20.3, and 42.6 on Humanity's Last Exam without tools against 36.8. On the same table Anthropic's Opus 5 leads most rows, including 51.8 on Terminal-Bench 4.0.
- DeepSeek also opened its Search and Fetch APIs to all users at $0.75 per 1,000 calls per endpoint, free through server-side tools in DeepSeek Harness, Claude Code and Codex until billing starts on 31 October 2026 at 16:00 UTC.
- Every figure here is DeepSeek's own and no independent evaluation has been published. The company does not say how the model differs architecturally from the version it replaces, and the card notes the Claude Code harness was used for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1.
Reflection AI publishes Beam's weights under Apache 2.0, 501B parameters with 23B active per token Company claim
- Reflection AI released Beam's weights on 16 October: a mixture-of-experts model with 501B parameters and 23B active per token, a 1M-token context window and an Apache 2.0 licence, in BF16 (about 1 TB), FP8 and NVFP4 (about 280 GB) checkpoints.
- The scores Reflection published for the BF16 checkpoint are SWE-Bench Verified 81.6%, SWE-Bench Pro v2-Hard 76.5%, Terminal Bench v2.1 80.1%, AIME 2026 97.7%, GPQA Diamond 90.6%, LiveCodeBench 88.3% and Humanity's Last Exam 36.0% without tools and 51.1% with tools.
- SiliconANGLE reports the model was introduced last week and that this release adds training recipes, serving-engine configurations, fused kernels and a drafter model for speculative decoding. It says Reflection trained Beam over two months using several thousand high-end graphics cards and credits custom reliability algorithms with practically zero cluster downtime.
- The benchmark table is presented as provided by the Reflection team, and no independent evaluation of the released weights has appeared. Reflection has not set pricing, and the Hugging Face post states no comparison of token use against other models.
Anthropic says an internal research model traced NIST's ECDSA curve seeds to lrand48 and Unix timestamps beneficialCompany claim
- Anthropic's Frontier Red Team says an employee asked an internal research model how the NIST curve seeds were generated, and that in under half an hour it found the origins of the P-192, P-224, P-256, P-384 and P-521 seeds. Anthropic writes that none of them were the results of any hash algorithm at all, but outputs of the C library's lrand48 generator seeded with a Unix timestamp.
- The prime-curve timestamps fall between May 1997 and April 1998; Anthropic gives 1997-05-18 14:45:34 for P-192, 1997-08-16 01:52:34 for P-256 and 1998-04-03 12:52:07 for P-224, and says the four binary-curve seeds B-233, B-283, B-409 and B-571 map to 17 April 1998 once the timestamps are halved. Recovering lrand48's 17 hidden bits meant trying 2^17 values, with state recovery by LLL lattice basis reduction.
- Filippo Valsorda, who posted the seeds bounty in 2023, writes that it has been claimed by ten people from Anthropic and that the seeds were generated with the POSIX-standard drand48 generator seeded with the then-current Unix timestamp. Anthropic says the $36,864 bounty will be donated to the Internet Security Research Group, which Valsorda identifies as the nonprofit that runs Let's Encrypt.
- Anthropic states that this is not a cryptographic break, new attack, or weakness in the ECDSA curves, and says it makes a deliberate backdoor less plausible. It also says the fifth binary seed B-163 remains unexplained, as do the base points for P-224, P-256, P-384 and P-521, and that a small number of viable candidate seeds were skipped for reasons that are not known.
Research & papers
Anthropic and Harvard report LLM-designed quantum circuits beating published designs on seven of eight tasks beneficialPreprintCompany claim
- The preprint, by Andi Gu, Alek H. Kemeny, Susanne F. Yelin and Mikhail D. Lukin, states that in seven of the eight tasks the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art, and that in the eighth the search discovers a new reversible architecture for prime-field modular inversion.
- Anthropic says Claude Mythos 5 was given plain-language specifications for eight primitives, wrote Python circuit designs and was scored by an automated evaluator on qubits multiplied by gates, with no human input between rounds. For elliptic-curve point addition, a 53-hour run produced a design using 12% more qubits than the smallest previous design but 17x fewer gates.
- For the decoder inside Decoded Quantum Interferometry, Anthropic reports gate count cut by a factor of about 12 against Google's previous construction, with the circuit reserving 19 qubits rather than a 665-qubit worst case, validated over 403 million simulated decodings with roughly one run in 10,000 failing. Anthropic notes most of the field's progress came from other groups, citing FeMoco estimates falling from about 50 trillion Toffoli gates in 2017 to about 340 million.
- The work is a preprint and the circuit figures are Anthropic's own. Anthropic says Claude was conservative, second-guessed apparent improvements and favoured well-known approaches, and that circuit design is one component of a fault-tolerant quantum computer rather than a complete system.
Epoch AI puts the best open-weight model's cyber capability 6-7 months behind the closed frontier Single source
- Epoch AI's data insight, by Jaeho Lee and Alexander Barry, puts Z.ai's GLM-5.3 at 160.10 on its Cyber ECI index (90% interval 155.83-166.97), the highest of any open-weight model. That sits above OpenAI's GPT-5.4, released 5 March 2026 at 157.04, and below Anthropic's Claude Mythos Preview April checkpoint at 164.82, which Epoch uses to place the open frontier roughly where the closed frontier stood 6-7 months earlier.
- The closed frontier is led by Claude Opus 5.5 at 171.14 and GPT-6 Astra at 171.03. The data is as of 8 October 2026, and the October update expanded the index from 56 to 76 models and from 20 to 35 cyber benchmarks, covering 324 model-benchmark results.
- Epoch states the limits plainly: GLM-5.3's score rests on three CAISI benchmark results and Opus 5.5's on four benchmarks reported by Anthropic, those two models share no benchmarks and their intervals overlap almost entirely, and CAISI did not name the US models it evaluated, so Epoch attributed those results to the models it judged most likely.
- The October refit lowered Claude Mythos 5 by 3.9 points and GPT-5.6 Sol by 4.7, so Epoch says earlier Cyber ECI values are not directly comparable. Only models with at least two cyber benchmark results are charted, so some recent open-weight models do not appear.
Redwood Research pipeline turned 94 real agent transcripts into 35 attacks that evaded its monitors mixedPreprintSingle source
- The preprint Screwtape, by Gosha Kozhevnikov and Tyler Tracy, describes a pipeline in which an LLM attacker edits real agent transcripts so that the agent causes a specified harm, such as leaking an API key or spawning an unmonitored agent. The edited actions are never executed, and an LLM verifier rejects edits that conflict with the original transcript or fall below a severity threshold.
- Run on 94 Claude Code and Codex transcripts from Redwood Research staff, the abstract reports that across 252 runs it produced 35 attacks that passed the verifier and evaded the authors' own monitors.
- Six of those 35, confirmed valid by human review, also evaded a held-out monitor provided by a frontier developer. The authors conclude a red-teamer without access to a developer's infrastructure can still find confirmed failures in that developer's monitor.
- The paper is a preprint and the arXiv title page lists Redwood Research only for Tracy, with no affiliation for Kozhevnikov. The abstract does not say which developer supplied the held-out monitor, and no independent replication exists.
Benchmark finds LLM judges accept 88.6% to 94.1% of tampered deployments producing identical outputs mixedPreprintSingle source
- The preprint, by Jianchang Su, Dawei Xiang, Yifan Zhang and Wei Zhang, introduces DIB-Bench, covering 7,560 scored examples and six tamper types across an evaluator ladder from rule-based scorers through LLM judges to GPT-5-nano.
- Mean acceptance of tampered outputs rises with evaluator strength, from 88.6% for rule-based evaluators to 94.1% for GPT-5-nano. The abstract summarises this as stronger evaluators accepting more tampered outputs.
- A trigger-conditional tamper that is invisible on the evaluation suite produces 48% to 100% harmful outputs, and that share rises with model capability. The authors give a PAC-style bound showing tampers with zero output shift cannot be detected from outputs at any sample size, while a manifest-hash identity channel detects them with zero samples.
- Behavioural fingerprinting identifies the model component at 76.7%, above chance, but covers only one of six stack components. The paper is a preprint and the arXiv abstract page does not display the authors' institutions.
Security, misuse & threat intelligence
Microsoft and NVIDIA publish a joint whitepaper on open-weight model risk with a five-category taxonomy beneficialCompany claimSingle source
- Microsoft and NVIDIA published Risk Management for Open-Weight Models in Enterprise on 16 October, described in Microsoft's post as the first in a planned series of joint resources. The post is by Michael Scovetta, Partner Security Assurance Manager at Microsoft.
- The paper sets out a shared risk taxonomy in five categories, Cybersecurity, Data Protection, Content, Human-AI Interaction and Bias, and is grounded in ISO standards and the NIST AI Risk Management Framework. Microsoft writes that weight openness alone does not determine an AI system's risk, which depends on the model, use case, deployment and safeguards.
- Microsoft says the paper maps evaluations to a selection of risks across those categories and is a starting point rather than a policy or checklist, focuses primarily on the model layer, and notes that enterprises operating their own model deployments may take on controls otherwise provided by a managed service.
- This is guidance from two vendors with commercial interests in enterprise model deployment, not an independent assessment, and the post reports no incident data, measured failure rates or evaluation results.
Police used an AI-generated image of a fictitious teenager to arrest a school worker, 404 Media reports mixedSingle source
- 404 Media, reporting with Court Watch on 16 October, says the Livonia Undercover Resistance to Exploitation Unit ran an operation in which FBI Task Force Officer McKailey Scott created a profile on the dating app Taimi using an AI-generated image of a teen-appearing male named Caleb, with no age listed.
- According to the affidavit as reported, Christopher Thomas Mertens, a high school maintenance worker, began messaging the persona in September, moved the conversation to WhatsApp, was told the persona was 15, sent sexually explicit messages and was arrested on arrival at the meeting place.
- The same affidavit describes investigators using Thomson Reuters CLEAR to look up his phone number, Clearview AI facial recognition to identify him from his photos, and Flock to track his car to the meeting spot, stacking a generated persona on top of commercial identification and surveillance tools.
- 404 Media says Livonia police did not respond to a request for comment and that the complaint does not name the school. The report does not list the specific charges, and no court has ruled on the evidence or on the use of a generated image in an undercover operation.
Benchmark on real criminal cases: unsafe LLM compliance rises from 28.9% to 66.4% under corrective framing harmfulPreprintSingle source
- The preprint, by Mohsinul Kabir of the University of Manchester and co-authors at the Islamic University of Technology, ELLIS Institute Finland, the University of Turku and Manchester, builds GUISE from criminal-assistance scenarios drawn from real cases organised under the UN Office on Drugs and Crime's International Classification of Crime for Statistical Purposes.
- The abstract reports mean unsafe compliance rising from 28.9% under direct requests to 66.4% under correction-seeking framing, peaking at 97.3%, with p < 0.05.
- Culturally adapted prompts drew 7 to 18 percentage points more unsafe compliance than literal translations across all four non-English languages tested, and leakage of Anglocentric register in unsafe responses rose with harmfulness, at rho = 0.46 to 0.63 across languages, all p < 0.001.
- The paper is a preprint marked under review, and the models evaluated are named in the paper HTML as Claude Haiku 4.5, DeepSeek V4 Pro, GPT-5.4 Mini, Gemma 4 31B, Llama 3.3 70B and Qwen3 32B. The authors do not report whether the vendors were notified, and no vendor has responded publicly.
Military, defense & geopolitics
China's draft export catalogue adds 40 technologies, 23 of them AI, with 14 prohibited and 26 restricted Update
- Global Times reports the draft revision of China's catalogue of technologies prohibited and restricted for export, issued jointly on 16 October by the Ministry of Commerce and the Ministry of Science and Technology, adds 40 technologies: 23 in AI, 12 in photovoltaics and 5 in biomedicine, split into 14 additions to the prohibition category and 26 to the restricted category.
- On the prohibited side are core technologies of AI large models trained with national key sensitive data, and AI agents serving key national information infrastructure, national defence and core government affairs. Restricted items include design technologies for 2.5-dimensional and 3-dimensional AI chips, 3-dimensional flash memory, AI information processing and voiceprint recognition.
- Technologies that have entered the public domain through open source, published academic papers or other means are outside the scope of management, and Global Times says automatic licensing applies to outbound licensing models that do not involve a transfer of ownership.
- Yesterday's edition reported the 23 AI technologies and the registration requirement taking effect on 16 October; new here are the three-category breakdown, the prohibited and restricted split, the named AI items and the open-source exemption. The comment period runs to 15 November and a final version has not been published, so none of this is yet in force as a final rule.
Taiwan to add advanced AI chips and servers to its export control list, economics minister says Single source
- Taiwan News reports Economics Minister Kung Ming-hsin said in a Bloomberg interview that Taiwan would add advanced semiconductors and AI servers to its export control list, tightening shipments to China and bringing its rules closer to US restrictions. He said unauthorised shipments of restricted equipment could then lead to criminal charges under the Foreign Trade Act.
- Kung said the changes are part of commitments under the Taiwan-US Agreement on Reciprocal Trade and that some controls would follow US rules, but did not say when the measures take effect.
- Taiwan already requires permits for exports of sensitive technologies, but Taiwan News, citing CNA, says advanced AI chips and servers are not specifically covered, which has made some unauthorised shipments difficult to prosecute. In August prosecutors indicted nine people over an alleged scheme to export AI servers to China involving 130 servers, 74 of them allegedly routed through other countries using false documents; the charges filed were breach of trust and document forgery rather than export offences.
- This rests on a single outlet's account of a Bloomberg interview, with no effective date, no list of covered items and no draft text published. Taiwan News describes the interview as taking place on Monday, which the article does not reconcile with its own publication date.
Acting Navy secretary says the Theodore Roosevelt strike group deployed on 8 October with an unmanned surface vessel Single source
- DefenseScoop reports acting Navy Secretary Hung Cao told an AUSA conference panel on Thursday: 'The Teddy Roosevelt Carrier Strike Group just deployed on October 8 with a medium-sized unmanned surface vessel. This is the first carrier strike group to ever deploy with a large-size unmanned surface vessel, and this is a force multiplier.' DefenseScoop flags the discrepancy between medium-sized and large-size in the same quote.
- The carrier is moving towards the Middle East after departing California late last month. Cao also said he had appointed Christopher Miller as the Navy's portfolio acquisition executive for robotic and autonomous systems, an organisation launched earlier this year.
- DefenseScoop sets the deployment against the Navy's 'Golden Fleet' plan to fuse AI and robotic platforms with manned ships, and notes the service merged its LUSV and MUSV programmes into the Modular Attack Surface Craft programme last year.
- Cao did not specify the vessel model, multiple Navy spokespeople did not clarify it, Pentagon officials referred questions to the Navy and a Leidos source referred DefenseScoop to the Navy. The report gives no autonomy or AI specifications for the vessel actually deployed.
Germany tells Boeing it will not buy the MQ-28 Ghost Bat, pushing its drone wingman to the early 2030s Single source
- Breaking Defense reports Germany's Ministry of Defence formally notified Boeing that it will not directly procure the MQ-28 Ghost Bat drone wingman. A Boeing spokesperson said the ministry had informed the company of the move and that Boeing stands ready to support Germany on next steps; Breaking Defense credits Politico with first reporting the decision.
- Germany had targeted 2029 to field a Collaborative Combat Aircraft, which would have made it the first European country to reach that milestone. Per Politico as cited, Berlin will now wait until the early 2030s, with aircraft tests starting in 2027.
- Remaining contenders named are Helsing's CA-1 Europa, which the firm says will be operational by 2029, the Airbus and Kratos U740-Valkyrie, which Airbus says is on track to deliver a combat-ready uncrewed CCA by 2029, and a possible General Atomics offering from its Gambit family.
- A German ministry spokesperson declined to explain the reasoning, citing procurement and public contracting law and the requirement that such matters first go to parliament. The report gives no detail on the autonomy or AI software in any of the competing aircraft.
Health, science & medicine
Google Research agents reach a 42.1% aptamer hit rate at 98.5% specificity after three wet-lab cycles beneficialCompany claimSingle source
- In a post by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, Google Research reports three design-test-iteration rounds against alpha-thrombin, each cycle taking 7 to 10 days. Round 1 produced 0% binders, with fixed primer overhangs disrupting folding in 48% of sequences; Round 2 reached a 26.0% hit rate after the agents added clamped flanking stems; Round 3 reached 42.1% at 98.5% specificity.
- Against myoglobin the agents went backwards before recovering: in Round 2, G-quadruplex motifs made up 63% of designs, specificity fell from 75.0% to 19.1% and the bead-corrected hit rate fell from 10.2% to 3.7%, because most of those binders were sticky and bound off-target proteins and bare beads. Round 3 restored specificity to 72.4% and reached a 33.5% on-target hit rate, which Google describes as roughly ninefold over Round 2.
- For de novo FAP binder design, Google says surgical epitope excision sped up the structure re-folding step by 8.6x and that at equal compute human-agent collaboration nearly doubled the in-silico hit rate of a scientist working alone, with top designs yielding single-digit nanomolar affinity binders. In a self-play campaign reconstructing blinded dye mixtures, its VERID agent improved accuracy more than fivefold across runs.
- All figures are Google's own and the three underlying papers are hosted as PDFs with no arXiv identifiers. Google does not state the baseline for the fivefold VERID figure, gives no exact affinity value for the FAP binders and gives no percentages for the human-agent comparison.
Multimodal AI beats InSilicoVA on cause of death by 8.5 to 15.9 points in 4,999 South African verbal autopsies beneficialPreprintSingle source
- The preprint, by Yue Chu, Zhenke Wu, Tyler H. McCormick, Zehang Richard Li and Samuel J. Clark, benchmarks language models on free-text narratives, machine learning on structured questionnaires and combinations of the two against InSilicoVA, using 4,999 verbal autopsy records for deaths at age 12 and older from a national sample of registered deaths in South Africa, with physician-assigned causes as the reference.
- Every supervised pipeline improved on InSilicoVA by 8.5 to 15.9 percentage points in top-1 accuracy and by 3.5 to 7.1 points in cause-specific mortality fraction accuracy. Feature fusion with cross-modal attention reached the highest CSMF accuracy of any pipeline at 95.7%.
- Decision-level fusion added 4.2 to 5.4 points of top-1 accuracy over single-modality models, mainly on the most common causes, while feature fusion with cross-modal attention added 6.6 points of balanced accuracy (95% CI 1.7-11.5) and 5.5 points of macro F1 (95% CI 0.6-10.3), mainly on less common causes. A zero-shot large language model reached the highest balanced accuracy of any pipeline, at 60.6%.
- The paper is a preprint. The authors report that paired differences from the better single-modality model ranged from -0.2 to +3.4 points with all intervals including zero, and write that external validation across settings and age groups will support broader implementation.
Audit of 2,550 conversations finds chatbot persistence after refused crisis referrals ranging from 0.01 to 9.94 mixedPreprintSingle source
- The preprint, by Subigya Nepal, Ruth Heo, Emma M Parrish and Colin A Depp, ran 2,550 scripted multi-turn conversations in September 2026. Each opened with a disclosure of suicide thoughts at one of 3 severity levels, answered every referral with up to 10 standardised refusals, then changed the subject.
- All 2,550 conversations included a referral, 2,542 of them (99.7%) in the first reply, but mean persistence ranged from 0.01 to 9.94 refusals across chatbots. Newer versions of all 5 chatbots tested in two versions persisted less than older versions, each with P < .05.
- Compared with passive ideation, a disclosure of intent was associated with 1.26 more refusals sustained (95% CI 1.06-1.46) and higher odds of a renewed referral after the topic change, at an odds ratio of 10.07 (95% CI 7.43-13.66). Of 8,818 refusal-phase replies urging outside help, only 173 (2.0%) asked about barriers to seeking it.
- Chatbots that persisted less in referral more often presented themselves as the user's support, at Spearman rho = -0.80, P < .001. The paper is a preprint and the abstract does not name which products scored at either end of the range; the authors argue safety evaluations should test what happens after a referral is refused, not only whether one is offered.
Policy, regulation & law
NIST, NSA and CISA publish ten voluntary practices for incidents during super intelligence training beneficialSingle source
- Published on 16 October by NIST's Center for Advancing Innovation and Standards for Super Intelligence, the National Security Agency's Super Intelligence Security Center and CISA, the document sets out ten voluntary key practices for frontier developers, aimed at incidents in which SI agents take significant uninstructed actions that undermine the operator's policies, controls or human oversight measures.
- The ten practices are organised into five stages of the incident management lifecycle: preparation, detection, response, recovery and reporting. Under reporting, the guidance asks organisations to voluntarily report cybersecurity incidents or near-misses arising from control incidents to CISA, and to share SI-specific information with NIST CAISSI and NSA SISC.
- NIST says the document builds on SP 800-61r3, on CAISSI's own experience running secure evaluations of frontier models and safeguards, on a recommendation in the 2025 SI Action Plan and on a May 2026 NIST workshop on SI incident management, with further work on agentic incident response planned.
- Everything in it is voluntary, with no reporting mandate and no enforcement mechanism. The guidance treats actions resulting from an adversarial prompt injection or an intentionally backdoored model as instructed actions, and so not the primary cause of a control incident, which leaves a category of real-world agent failures outside its scope.
Anthropic confirms it is circulating a draft federal AI bill with $1 billion and 50-casualty thresholds Single sourceCompany claim
- Anthropic confirmed to AFP on Friday that it is circulating a draft bill among US lawmakers. Under the draft, first reported by Politico citing three people with knowledge of the efforts, Washington could block a model's release or use only where there is a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or of causing more than $1 billion in damages, and only with a judge's approval.
- The risk must stem from the creation or release of a biological weapon, the model evading its developer's controls, or a cyberattack. Independent evaluators registering with the Commerce Department would periodically assess the lab and its frontier models, may be paid by the labs provided payment is not tied to the outcome, and must report an imminent enumerated risk to Commerce within 72 hours of completing an assessment.
- Reuters reports the Commerce Department could ask a federal judge for the District of Columbia to order a suspension, while the language explicitly prohibits any other US department, agency or officer from suspending development, deployment or internal use. AFP reports the bill would also override certain state laws, including California's, and notes Anthropic wrote in June that absent a strong federal law, states should retain full authority to legislate.
- An Anthropic spokesperson told AFP the proposal grants the federal government the ability to block or deter dangerous deployments of the most powerful models, and told Reuters it represents proposals in Anthropic's Advanced AI Framework. The draft has not been introduced, Congress is out of session until after next month's midterms, the Commerce Department did not immediately respond to Reuters, and no text has been published, so every provision above is reported rather than read.
Unsealed filing says Microsoft told staff to keep the OpenAI crawler project codenamed Project Mango secret Single sourceUpdate
- Bloomberg Law reports that a court document unsealed on Friday says Microsoft ordered employees to keep secret a website-scraping project that harvested AI training data for OpenAI. In the effort codenamed Project Mango, the publishers' partly unsealed motion for summary judgment says OpenAI paid Microsoft to develop and operate a crawler to copy webpage content, which was then used to train models including GPT-4o.
- That data allegedly included copies of more than 160,000 of the publishers' works. How much OpenAI paid Microsoft and the number of webpages scraped remain under seal, as do how many training licensing deals the companies struck and how much they paid.
- The filing says the publishers, including the New York Times, estimate ChatGPT fields 47 million prompts a day seeking news content, derived by applying an OpenAI expert witness's figure of 1.88% of all ChatGPT messages in July 2025 seeking news to the 2.5 billion prompts a day the chatbot receives. The document also discloses that OpenAI gave Microsoft its full set of GPT-3 training data as part of self-described horse trading.
- The case is In Re OpenAI, Inc. Copyright Infringement Litigation, S.D.N.Y. No. 1:25-md-03143, with the motion partially unsealed 10/16/26. These are the plaintiffs' submissions on cross motions for summary judgment, not court findings; Microsoft had argued for continued redaction and withdrew some requests on 9 October. Bloomberg Industry Group, which operates Bloomberg Law, has moved to intervene to oppose the sealing.
Apple updates App Review Guidelines to cover AI-generated nonconsensual intimate imagery and bait-and-switch apps beneficialSingle source
- In a developer post dated 16 October, Apple says guideline 1.2 now explicitly covers nonconsensual intimate imagery, including imagery generated using AI tools, as content apps must proactively prevent and remove. Developers who fail to implement adequate safeguards, or do not remedy violations after notice, risk violating the Developer Code of Conduct and removal from the App Store.
- Guideline 2.5.19 clarifies the prohibition on apps that misrepresent what they do during review and then change after approval, which Apple describes as the bait-and-switch tactic used by nudification apps that only reveal harmful capabilities once approved. Offending apps face removal from the App Store and from the Apple Developer Program.
- Apple also updated Section 5 to state that apps facilitating human trafficking or the exploitation of children, including child sexual abuse material, will be removed from the App Store and blocked from alternative distribution. Apple says that last year it prevented over 150K apps with safety issues from reaching customers and terminated over 193K developer accounts for fraud and abuse.
- These are rule changes and stated enforcement posture, not measured outcomes: Apple gives no figure for nudification apps removed, no timeline for re-review of apps already on the store, and no detection method for AI-generated imagery.
Compute, chips & infrastructure
Anthropic is offering Blackstone a yield of more than 10% on $9 billion of chip-lease debt, Reuters reports Single source
- Reuters, in a wire story dated 16 October and citing the Financial Times, reports that Anthropic is offering Blackstone a yield of more than 10% to finance its latest chip deal, and that Blackstone funds would invest $9 billion in debt backed by Anthropic's five-year lease payments on the chips.
- The FT's account, as relayed by Reuters, attributes the terms to people familiar with the matter. Reuters states plainly that it could not immediately verify the report.
- Anthropic has not published the terms, the package has not closed and neither Anthropic nor Blackstone is quoted. Reuters gives no total size for the financing, no chip vendor and no closing date.
- Yesterday's edition reported investors discussing an Anthropic IPO valuation near $1.5 trillion, down from the roughly $2 trillion previously expected; the cost of this debt is a separate data point on the same question and is not independently confirmed.
NVIDIA says RTX Spark laptops ship from six makers with up to 1 petaflop FP4 and 128GB unified memory Company claim
- In a release dated 16 October, NVIDIA says RTX Spark laptops are available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, with additional systems to follow, starting at $2,599. NVIDIA states the platform delivers up to 1 petaflop of AI compute, specified in the body as up to 1 petaflop of FP4 AI performance, and up to 128GB of unified memory.
- The superchip combines an NVIDIA Blackwell RTX GPU with an NVIDIA Grace CPU, with MediaTek collaborating on the custom CPU design. NVIDIA says Nemotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter model optimised for tool calling and multistep tasks, runs on supported configurations.
- NVIDIA names day-one agent support from GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, alongside its OpenShell secure runtime and Microsoft Execution Containers for agent permissions, and says NVIDIA PAIR lets an RTX Spark machine serve inference to other Windows, Linux and macOS machines on a home network.
- All performance figures are NVIDIA's own and no independent benchmarks have been published. The release gives no measured inference throughput for the Nemotron model on the hardware, and says pricing and configurations vary by region.
Digital Realty plans a 1GW data centre campus on 476 acres in Kokomo, Indiana, built out by 2033 Single source
- Data Center Dynamics reports on 16 October that Digital Realty has announced plans for three single-storey data centres on 476 acres in Kokomo, Indiana, about 50 miles north of Indianapolis, on greenfield land in the city's Industrial Park North.
- DCD says the campus could total around 1GW of IT capacity across some three million sq ft, equal to 278,710 sqm, and could be fully built out by 2033. Digital Realty does not currently operate any data centres in Indiana.
- The city of Kokomo first announced Digital Realty's interest earlier this month, has said the site will use a closed-loop cooling system to reduce water use, and has said there will be no tax abatements agreed with any company that develops the site.
- The capacity and completion date are framed as possibilities rather than commitments, with no power agreement, construction start or customer named, and no filed planning application reported.
UK rejects an AI Growth Zone bid for a 1.2GW data centre powered by Drax's biomass plant Single source
- Data Center Dynamics reports the UK government rejected a bid by North Yorkshire Council, the University of York and Drax to use the bioenergy plant near Selby to power an AI data centre, submitted through the AI Growth Zone scheme that offers state funding and other benefits for large-scale AI data centres.
- The proposal envisaged a 1.2GW data centre powered by the 2.58GW plant, the UK's single largest energy generator, which the applicants described as the world's largest carbon negative data centre. They claimed carbon capture in partnership with Mitsubishi Heavy Industries would capture at least eight million tonnes of CO2 a year; DCD cites Ember data showing Drax emitted 14.1 million tonnes of CO2 equivalent in 2025.
- Per local reports by the York Press cited by DCD, the bid failed to meet selection criteria including credible plans for energy delivery, environmental considerations and local economic benefit. A Drax spokesperson told the York Press it had never relied on the bid's success.
- Drax has pivoted to a 100MW front-of-meter data centre drawing from the national grid, which it says it plans to submit to planning authorities in the coming months. The government's reasoning reaches DCD second-hand through the York Press rather than from a published decision.
AI cloud Modal raises $750 million at a $15.75bn post-money valuation, five months after its Series C Company claimSingle source
- Data Center Dynamics reports on 16 October that Modal secured $750 million in a Series D that the company says gave it a $15.75bn post-money valuation, led by Accel with Kleiner Perkins and existing investors Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures.
- The round comes five months after a $355m Series C at a $4.65bn valuation, which itself was more than four times the $1.1bn valuation of its $80m Series B in September 2025.
- Chief executive Erik Bernhardsson wrote on LinkedIn: 'Since then, usage has tripled and we've grown more than 15x year-over-year. We recently crossed $1 billion in annualized run rate with no sign of slowing down.'
- Modal does not own or operate data centres, working instead with 13 cloud partners, and says users can scale from zero to 1,000 GPUs in minutes by pooling capacity across hundreds of data centres. The valuation, growth multiple and run-rate figure are all the company's own and are not audited or independently verified.
Deployment & impact
Anthropic says the junior share of its technical hires fell from 27% to 9% since early 2025 mixedCompany claim
- In a report published 16 October by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, Anthropic says the junior share of employees in its technical departments slipped from 21% to 14% since early 2025, and that the junior share of technical hires fell even faster, from 27% to 9%. Technical departments' share of total employment fell from 61% to 56% over the last year while the company tripled in size.
- Combined autonomous and interactive agent hours passed human work hours for the first time in March 2026, and by August interactive agent hours outnumbered employees' 6 to 1, counting employee-directed sessions and assuming 40-hour human weeks. Code commits per engineer ran at about 5x their velocity six months earlier, and coding-agent sessions among employees with no code commits at baseline rose from 2 a month in August 2025 to 90 a month in May 2026.
- Other internal figures: an engineer-to-product-manager ratio of about 17:1, standing agent instructions growing from a few hundred in July 2025 to 8,500 a year later, about 80% of interactive Claude session time spent waiting for human input, and meta-work rising from 10% to 22% of interactive sessions between autumn 2025 and spring 2026.
- The productivity comparison follows 192 engineers who adopted coding agents against 74 who did not, and Anthropic states that adoption was not random, that employees who moved to autonomous agents were already on a steeply rising output trajectory, and that more code does not mean better code. Quality proxies such as code deleted or reverted showed no difference, but the authors note AI errors may be harder to detect. The authors also caution that Anthropic is not representative of the broader labour market.
Anthropic survey finds 91% of its staff and 39% of US adults expect a Nobel-level AI discovery by 2030 Company claim
- Anthropic reports that 91% of its staff expect AI to be able to make a Nobel-level discovery by 2030, against 39% of US adults. The US figures come from a Morning Consult survey of 10,980 adults fielded 11 to 23 August 2026 and weighted to be nationally representative; the staff figures come from an unweighted internal poll of 1,334 respondents run 9 to 12 August 2026.
- Staff put the share of their own current work hours AI could do by 2030 at 75% on average, against 28% for US workers, and more than half of staff expect 80% or more, against 5% of US workers. Depending on the task, 19% to 40% of US adults say AI will never be able to do it, against at most 2% of staff.
- Translated into macroeconomic terms, the median US adult's answers imply 5.4% GDP growth in 2030 and the median staff member's imply 17.1%. Starting from a stated 4.2% unemployment rate, US workers who expect 80% or more of their work to be AI-doable forecast 12.2% unemployment in 2030.
- Anthropic says the staff figures are the personal expectations of staff members in an internal poll and not a company forecast or plan. The US sample was drawn from opt-in online panellists, so it excludes adults without internet access; range answers were converted to midpoints; the GDP figures use only the 3,259 US adults and 1,093 staff who answered every key question; and a September follow-up of 7,613 US adults implied 5.6% growth, but the report uses the August results.
Replit's chief executive says it has stopped working with Anthropic; insiders say the partnership stands Single source
- Speaking at TechCrunch Disrupt, Replit chief executive and co-founder Amjad Masad said the company is no longer working with Anthropic: 'It was a decision based on both cost, partnership, and performance,' adding 'I think ChatGPT is better now' and saying Replit's public benchmark shows Sol and Astra perform better on building vibe-coded applications from scratch. Replit announced a partnership with OpenAI in August.
- People familiar with the Anthropic relationship rejected that account, with one telling TechCrunch the partnership was still in place. TechCrunch notes a late-September Replit blog post cited Claude Fable 5 and Fable 5.1 running in Replit Agent's production environment, and that a Replit product employee promoted Anthropic's Opus 5.5 on X around the same time.
- On the business, Masad said enterprise was about 2% of revenue last year, that about 15% of revenue now comes from enterprise contracts and that more than 50% of revenue is B2B. He did not share annual recurring revenue.
- The two accounts are in direct conflict and only one outlet has the story. Replit had not responded to TechCrunch's request for clarification at publication, and Anthropic is not quoted.
Vercel reports open-weight models took 71.5% of gateway tokens in September as Anthropic's spend share fell Company claim
- Vercel's AI Gateway Production Index, published 15 October and reported by TechCrunch on 16 October, says open-weight models processed 71.5% of gateway tokens in September for 23% of spend, with DeepSeek alone taking 58% of tokens for 6.6% of spend. Anthropic's share of total gateway spend fell from 63.7% to 43.9%.
- TypeSafe AI's Jev reached 35.3% of paid teams in its first two weeks, which TechCrunch reports is the fastest adoption of any model in its first two weeks on Vercel's data, roughly double the reach of the second most popular model.
- September gateway spend was about 13 times December's, an increase of 1,180%, while the average price per token fell 47%. Nearly 70% of September spend on Chinese-origin models went through US-based providers, up from less than 10% in December. Nearly one in two paid teams changed their most-used model and almost one in three changed their most-used lab, an index record.
- This is one platform's own telemetry, not an industry measurement, and TechCrunch notes the gateway data may undercount labs such as OpenAI and Anthropic that sell their own APIs directly.