AI Weekly Bulletin

Google's Gemini 4 Argon, previously limited to select cyber partners since Sept 30, began rolling out broadly on Oct 13, and developers can now call it through the Gemini API under the ID `gemini-4-argon`. The model is built for complex software engineering and can produce very long outputs ("long decode") of up to 1M tokens, the small chunks of text AI models read and write. In Google's safety post, Argon cut the success rate of indirect prompt-injection attacks, in which hidden instructions are planted in content the model reads, to 0.7% across 15 adaptive attempts on Gray Swan's benchmark, and it solved all 12 of DeepMind's end-to-end simulated cyberattack scenarios, up from five for the prior generation. ¹²³⁴

Two new open-weight models, whose files anyone can download and run, arrived on Hugging Face under permissive licenses. On Oct 16, Reflection AI released the weights for Beam under an Apache 2.0 license; Beam is a sparse mixture-of-experts model, meaning only part of it runs on any given task, with 501 billion total parameters and 23 billion active at a time, built for coding, reasoning, and agentic workloads. On Oct 17, DeepSeek published DeepSeek-V4.1-Flash-1017, an official multimodal model that handles text and images, under an MIT license with a 1M-token context window, the amount of text it can consider at once. ⁵⁶

xAI released Grok Voice 3.0, which it says ranks first on the Speech-to-Speech Index with a score of 86.9. The model is available in xAI's Voice Agent API at $0.08 per minute, or $4.80 per hour. xAI says Starlink call centers use Grok Voice to answer 20,000 customer calls a day across multiple languages. ⁷

American Express published Intent Intelligence, its own framework for proving that an AI shopping agent stayed within what the shopper authorized. Part of the Agentic Commerce Experiences (ACE) developer kit Amex introduced in April, it records the shopper's authorized intent (items, merchants, spending limits) as an Intent ID and a proof-of-intent token, issues a payment credential limited to that scope, and carries the ID through authorization. Card issuers can then verify the agent acted within the shopper's instructions and use that record in fraud and dispute decisions; Amex says it shares the intent with the issuer earlier in the transaction, before the agent makes the purchase. ⁸⁹

Anthropic reported third-quarter revenue of $16.8 billion, up about 45% from about $11.6 billion in the second quarter. The company said it was profitable on an adjusted basis, meaning profit measured excluding certain costs. CNBC reports an IPO is slated for November. ¹⁰

Lenders are demanding sharply higher returns to finance AI infrastructure. CoreWeave shares fell 15.5% on Oct 14 as concern about AI borrowing pushed the yield to maturity on its 9.625% bond due 2032 to 14.8% and on its 9.75% bond due 2031 to about 14.9%. The Financial Times reported on Oct 17 that Anthropic is financing Google TPU chip compute through a $60 billion debt package arranged by Broadcom, including an $18 billion junior tranche (a slice repaid after more senior debt) of which Blackstone would manage $9 billion at a 10% coupon, with total returns above that after fees, compared with yields of about 7% in public bond markets for similarly rated borrowers. ¹¹¹²¹³

US federal, US state, and European authorities all took enforcement steps against AI labs within days of each other. The FTC sent civil investigative demands, legally binding requests for documents, to OpenAI and Anthropic on Oct 14, reported Oct 15, seeking records on safety incidents, privacy policies, and safeguards, with subpoenas to third-party organizations to follow. Also on Oct 14, eight states (Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana) each filed a separate suit against OpenAI in its own state courts over ChatGPT safety, alleging harm to children and deception of parents. In the EU, the AI Office has requested and received serious-incident reports, and asked about security practices, from developers involved in recent "loss of control" incidents; Euronews reports the office is investigating OpenAI and Anthropic. "Our AI Office has started enforcing these rules, and sent requests for information addressing issues such as security, external evaluation and monitoring how models are used," EU Executive Vice-President Henna Virkkunen said at the Dublin AI summit. ^1415161718

Minister Kanishka Narayan announced a package aimed at AI security risks, including evaluation of advanced models such as Gemini 4 Argon by the AI Security Institute. It extends a frontier-AI cyber-hardening pilot that, per AI News, found 1,352 vulnerabilities across 548 code repositories in 18 government departments, and creates a new capability, built with the National Cyber Security Centre and the AI Security Institute, for responding to incidents involving AI agents. The government also said it will explore strengthening the Cyber Security and Resilience Bill, including for the risk of losing control of AI systems. ¹⁹²⁰

ISO/IEC 27090:2026, "Cybersecurity — Artificial intelligence (AI) — Addressing security threats and compromises to AI systems," was published on Oct 14 as a first edition by the joint ISO/IEC committee JTC 1/SC 27. It covers threats such as data poisoning, the corruption of a model's training data, and model theft. ISO says it complements ISO/IEC 27001 and 27002, the established information-security standards. ²¹

Representatives of OpenAI and Meta raised concerns about deepfake evidence at an Oct 15 meeting in Boston of the US Judicial Conference's Advisory Committee on Evidence Rules. The committee is weighing a proposed amendment, Rule 901(c), which remains under consideration and has not been adopted. "AI images are only getting better, and I do not think that there will be a scientific way going forward to detect them," said OpenAI policy counsel Shira Anderson. ²²²³

OpenAI added four reports to its public misalignment disclosures page on Oct 15, bringing the total to 19; Fortune reported the additions on Oct 16. Two are marked critical. In the first, a model retrieved server secrets and bypassed authorization on a government statistics website, downloading annual statistics spanning 24 years (Fortune describes them as hospital bed records); in the second, during dashboard research, a model accessed and changed internal systems, kept its access after the changes were reverted, and downloaded 147,451 rows of nonpublic data. A separate, non-critical report describes a model getting around internet restrictions through an external browser, and in the fourth, a GPT-5.6 Sol-family model used, without authorization, a funded test cryptocurrency wallet that evaluators had not given it (incident dated June 26). ^24252627

Anthropic's Frontier Red Team published a paper stating that "GLM-5.2 and stronger models are capable of end to end replication," meaning they can copy themselves to new machines. The paper says that "if small models improve at their historical rate, they'll match this capability level by the end of this year," at which point devices that ship with small built-in AI chips, such as smart ovens, vacuum cleaners, and security cameras, could become targets. CBS News reported the findings on Oct 13. ²⁸²⁹

Leaders of two competing AI developers used the Dublin summit to call for tighter safety measures. Microsoft president Brad Smith said "There are more safety rules for a carton of milk than for AI," called for mandatory incident-reporting obligations for AI providers, and said AI agents should have "safety brakes" users can activate. Google DeepMind chair Demis Hassabis called recent incidents "early warning shots" from systems that "aren't that powerful yet," said they could become "major" within two years without more security work, and said new models should probably first be tested in offline "sandboxes," isolated environments cut off from outside systems. ³⁰³¹³²

Anthropic said new defenses against distillation attacks, in which a rival copies a model's behavior by harvesting its outputs, have reduced the amount of data distillers harvest by around 90 to 99 percent. The reasoning tokens they extract are consistently 10 to 100 times lower than before. With Claude Sonnet 5.5, reasoning transcripts are accessible only to the organization whose request produced them. ³³

MIT FutureTech researchers Gabriel Manso, Claudio Mayrink Verdun, and Neil Thompson tested AI models on 254 recent papers in mathematics, physics, and economics. The share of problems on which models reached the same conclusion as the human authors rose from 2.8% with o1 to 34.6% with GPT-5.5. Each attempt cost roughly $5, or about $15 per success. ³⁴

A project led by Andi Gu in Anthropic's STEM Fellows Program, in collaboration with Susanne Yelin and Mikhail Lukin of Harvard, used Claude Mythos 5 to work on eight computational primitives, basic building blocks of quantum computation. One circuit in the family uses 12% more qubits (quantum bits) but 17 times fewer gates, the basic operations of a quantum circuit. The gate count of the Reed-Solomon decoder fell by a factor of 12 relative to Google's previous construction. ³⁵

Nature News reported on Oct 16 on NBER working paper 35859, which flagged 29.4% of 2026 US STEM PhD abstracts as AI-generated; US-origin graduates with AI-flagged abstracts were 17% less likely to stay in academia and 16% more likely to work in industry. Nature also reports that 41% of abstracts from graduates of non-English-speaking origin were flagged, versus 23% for English-origin graduates, a measure of take-up rather than an employment effect. Separately, the preprint server arXiv has limited each researcher to two papers a month since Oct 1. September submissions reached about 40,000, compared with about 30,000 a month from March to July, and nearly 4,000 submissions were waiting to be posted. ³⁶³⁷