The AI Wire ●

President Trump announced plans to appoint a new AI czar and create an 'AI Force' in a series of Truth Social posts, vowing to keep the US ahead of China in AI without slowing the industry. He dismissed safety concerns as a 'hoax,' writing that the administration 'will not in any way hinder or stifle the Growth of this incredible Industry,' and said it would address AI harms through existing criminal and civil law rather than new sector-specific regulation.

The announcement lands as a direct counter-framing to the month's safety cascade. It came within days of all four frontier lab chiefs publicly agreeing to pace development and accept embedded evaluators, plus California's first-in-the-nation AI safeguards bill. Trump also framed industry calls to slow down as a 'Market Fixing' ploy by figures like Anthropic's Dario Amodei and OpenAI's Sam Altman.

Concrete details are thin so far. As of mid-September there is no executive order establishing the AI Force, no legislation, no named czar — the role was previously held by David Sacks, who stepped down in the spring — and no stated budget, host department or remit. Axios reports the AI Force is modeled on the Space Force; with that much still undefined, this one stays 'announced' until the machinery shows up.

Researchers at AIR disclosed Plugin4Shell, a zero-click remote code execution vulnerability shared by four major AI coding agents — Claude Code, OpenAI's Codex, GitHub Copilot and Google's Gemini CLI. All four pin plugins to a reviewed commit hash but check out the pinned commit without verifying the checkout landed there, letting an attacker swap in malicious code while the pin still looks intact. AIR calls it the first supply chain vulnerability of the AI agent ecosystem.

The flaw is zero-click because the same git checkout re-runs on background auto-update: when a marketplace bumps the pinned SHA, the swap reaches already-installed plugins with no user action. AIR says it previously demonstrated plugins spreading to more than 26,000 agents and found 925 hijacked skills in active use reaching 134,000 agents, proving the takeover chain end to end.

Vendor response was uneven. Anthropic patched Claude Code in version 2.1.179 and OpenAI patched Codex in 0.146.0. Microsoft has not shipped a fix for Copilot, leaving its users with no patch. Google deprecated Gemini CLI entirely rather than patching it, meaning existing installs stay exposed indefinitely; its advice is to move to its newer agent, Antigravity.

OpenAI announced Astra for Law on September 18, a legal-focused version of GPT-6 Astra for lawyers and legal-tech companies. It pairs the model with a dedicated legal search index spanning more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, plus custom instructions for legal analysis and 26 ecosystem plugins connecting specialist legal tools.

OpenAI says it hit 54% correctness on 200 Vals AI Legal Research Bench questions versus 38.7% for plain GPT-6 Astra with web search. Initially available to selected US law firms through Trusted Access — announced, not broadly released, consistent with the site's restricted-tier convention.

Leakers claim a first Gemini 4 Pro checkpoint exists and point to October 2026 for launch, with Google skipping the delayed Gemini 3.5 Pro entirely. The Arena-disguise theory gained detail Sept 21: Dataconomy reports a 'gemini-3.8-flash' entry carrying the internal codename Argon, with leaked developer-shared benchmarks of 88% on DeepSWE v1.1, 95.3% on Terminal-bench 2.1, and 86.8% on OSWorld-2.0 — claimed to beat GPT-6 Astra and Claude Fable 5.1.

HTX Insights calls it a 'sneaky' Arena-only launch. Still no Google confirmation, and one leak-tracking site notes the whole chain traces to a single viral X post. Confirmed: Google says Gemini 4 pre-training began in July.

UPDATE 2026-09-21: a Dataconomy report claims developers are benchmarking a suspected Gemini 4 Pro under disguise (possibly the 'gemini-3.8-flash' label) on Arena: 88% DeepSWE v1.1, 95.3% Terminal-bench 2.1, 86.8% OSWorld-2.0, with a reported 10M-token input / 256K-token output window and $2.25/$11.25 per million pricing. Google has not acknowledged any Arena testing or confirmed specs, timing or the model; all numbers are attributed to unnamed developers.

Community estimates still point to an October public release, also unconfirmed. UPDATE 2026-09-21: a Nexscope breakdown of the circulating demo clips (a butterfly with explorable parts, a watch that separates into layers, character interfaces, SVG/3D/interactive generation) frames the Gemini 4 Pro vs GPT-6 Astra showdown around creative and interactive output, but its author stresses the demo clips do not establish a launch date, API identifier or price — the checked Google catalog still has no Gemini 4 Pro entry.

UPDATE 2026-09-22: a new data point — an AI-news roundup reported 'a mysterious Gemini Pro checkpoint' appearing in testing, completing a visual benchmark faster than OpenAI's GPT-6 Astra. Consistent with the disguised-testing half of this rumor, but single-source and unverified.

UPDATE 2026-09-22: the story keeps circulating on the same September 17 Arena evidence with no new corroboration. One breakdown (getneotiler) notes the leaked benchmark table’s Astra reference scores don’t match Astra’s published numbers, and calls the whole table possibly fabricated — Google still hasn’t commented on the Arena listing or confirmed anything about Gemini 4 Pro.

UPDATE 2026-09-23: still no official Google signal. New secondary chatter only: Van Data Team’s FAQ (updated a day ago) summarizes leaker claims of a first checkpoint live with an October target, and a Polymarket market has “Oct 12–18” as the frontrunner at 39% (19% on no release by Nov 1). Markets and leaker roundups are estimates, not confirmation — this stays unresolved.

UPDATE 2026-09-24: a Sept 24 tech-industry roundup cites The Information on Google's position — its flagship lag has been the opening for rivals in coding and agents, and an earlier Gemini 4 ship would reopen the frontier-model race before 2027 budgets lock in. Still no official word from Google on timing or specs; this stays unresolved.

CONFIRMED 2026-09-25: Google DeepMind's new SVP, Koray Kavukcuoglu (took over day-to-day operations in August), said on-record Wednesday at The Information's AI Agenda Live Summit that Gemini 4 is in early post-training and will arrive 'much earlier' than the end of the year. That is the most concrete timing signal from a Google leader yet, and it corroborates the October-target half of this rumor — while the Arena-disguise testing and leaked benchmark claims remain unattributed to Google.

CONFIRMED 2026-09-30: Google announced the model as "Gemini 4 Argon," confirming the October timing — and, notably, the codename "Argon," which the Sept 21 Arena-leak reporting had already attached to the suspected checkpoint. See the Gemini 4 Argon announcement entry.

The Wall Street Journal reported on September 18 that Anthropic is moving ahead with a planned November IPO targeting a valuation of roughly $2 trillion and a raise of up to $100 billion — which would rank among the largest listings in market history. The timeline slipped from October to November so the company can present fresh third-quarter financials to investors before locking in pricing.

The valuation would reflect explosive growth: annualized revenue is reportedly above $100–110 billion this year, up from $65 billion disclosed in July. The Financial Times cautions the figure is not formally fixed inside the company, and none of it comes from audited filings yet.

No public S-1 has been filed and neither Anthropic nor its lead investors have confirmed the timeline. By contrast, Sam Altman has confirmed OpenAI will not go public in 2026, with 2027 the earliest possibility.

On October 1, Bloomberg reported that formal IPO marketing could begin as early as the week of November 9, setting up a debut before the November 26 Thanksgiving holiday. Prospective investors reportedly value the company at $1.8 trillion to $2 trillion despite its losses, with 2025 revenue around $4.6 billion and a net loss of nearly $42 billion.

On October 2, Bloomberg reported further timeline detail: Anthropic plans an Investor Day at its San Francisco headquarters on October 14, aims to begin IPO marketing as early as the week of November 9, and wants to go public before or around the November 26 Thanksgiving holiday — a debut that could be the largest IPO in history. An Investor Day source said discussions on the listing are ongoing and the timeline could still shift.

On October 3, AI Startups News reported the IPO roadshow is set to begin in mid-October with formal marketing starting the week of November 9, targeting a $1.8–$2 trillion valuation ahead of the November 26 Thanksgiving holiday — a debut that could be the largest IPO in history. The report also noted Anthropic has posted two consecutive quarters of positive adjusted operating income and has ended enterprise discounts once usage caps are reached. OpenAI, meanwhile, is reportedly near a $30B pre-IPO raise at a $1.4T+ valuation but plans to hold its own IPO no earlier than 2027.

Anthropic has set up a physical wet lab in the San Francisco Bay Area, Reuters reported on September 18. The company's head of life sciences, Eric Kauderer-Abrams, confirmed the facility, an unusual move for a company best known for large language models.

The rationale is a feedback loop between prediction and reality: Claude can design candidate molecules and analyze results, but the hard proof comes when a physical experiment succeeds or fails. Anthropic had already shown why physical validation matters — in August it said external evaluators produced and tested Claude-designed protein binders against 15 targets, confirming binders against 14 of them.

Anthropic drew a careful commercial line around the lab: it will stop short of running clinical trials, and a spokesperson told reporters the lab is "not for drug discovery specifically." The positioning keeps Anthropic selling tools to drugmakers rather than competing with them.

Anthropic has named Accenture as the first "embedded evaluator" under CEO Dario Amodei's "Pace the Frontier" plan. Each company will invest at least $1 billion over five years to build an in-house evaluation team that tests Anthropic's models from inside the company.

The unusual part is the access: instead of testing finished models from the outside, embedded evaluators get "access comparable to an employee's" — observing training runs, monitoring deployment decisions, red-teaming models, running alignment assessments and testing safeguards. Anthropic says this makes accountability "more verifiable" without reducing its own responsibility for model safety.

The choice raised eyebrows in the safety community: Accenture's Faculty unit is a commercial consultancy, not a nonprofit like METR, and more than 100 researchers — including Geoffrey Hinton — publicly demanded evaluator independence from commercial ties the same day. Anthropic says the deal is non-exclusive and more evaluators will be announced.

Anthropic announced on Friday that Accenture will become its first 'embedded evaluator,' the first concrete step in executing CEO Dario Amodei's September 12 proposal to slow the pace of frontier AI development. Accenture's AI unit Faculty will place personnel inside Anthropic with employee-like access to evaluate and red-team models, assess alignment, and test safeguards across the training, development and deployment pipeline.

Both companies expect to invest at least $1 billion each over the next five years to build out the capability. Anthropic says it will fund Accenture's work directly in the near term because no established financing mechanism for independent evaluation exists yet. The arrangement is non-exclusive — Anthropic expects to name additional evaluators in the coming weeks.

The market noticed: Accenture shares jumped 8.8% after the announcement. The pact follows Anthropic's disclosure that Claude is leading 26% of its own model R&D — the recursive-self-improvement dynamic Amodei's slowdown proposal is built around.

OpenAI on September 17 launched Astra for Law, a version of GPT-6 Astra configured for legal work, calling it the beginning of a long-term investment in law. The product pairs the Astra model with a dedicated legal search index covering more than 230 million URLs of US case law, statutes, regulations and court rules, plus custom instructions for legal analysis and writing.

Distribution starts with API customers including Harvey and Legora, followed by a 'Trusted Access' program aimed at Am Law 200 firms, with a new zero data retention policy for the API and ChatGPT Enterprise use excluded from human review. OpenAI said it worked with firms including Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell Lipton on testing.

The launch also brought a plugin marketplace (26 vendor plugins including Thomson Reuters, Harvey, Legora and iManage, plus 47 custom skills) and the general availability of ChatGPT for Word. In OpenAI's own benchmark on 200 Vals AI legal-research questions, Astra for Law scored 54.0% correct versus 38.7% for the base model on web search.

Anthropic announced on September 17 a redesigned Projects experience for Claude Code, in beta for a select group of Pro and Max subscribers who use cloud sessions. The update replaces the static reference-file folder with a coordinator that scopes a request, delegates work to parallel agent threads (each a full Claude Code cloud session on its own branch), reviews outputs and assembles results.

Threads share a memory layer that carries decisions and context across sessions, plus a library of files and generated artifacts. Users can steer progress from a phone, and work continues after they step away. Anthropic says access widens over the coming week, with the rest of Claude and Team/Enterprise plans to follow. (Filed September 21 as backfill.)

OpenAI began testing 'Sponsored Agents' in ChatGPT on September 16 — a new ad format where tapping an ad opens a clearly labeled conversation with an advertiser-run agent instead of redirecting to a browser tab. Users can ask follow-up questions about a product and click through to the business's site when ready. Wayfair and Angi are among the first participants, alongside Newegg, Best Buy, Lowe's, and VistaPrint.

OpenAI says the agents stay clearly labeled and separate from ChatGPT's own responses and the user's original session. The pilot is limited to select US advertisers, and OpenAI hasn't disclosed how often a sponsored agent surfaces or what triggers one.

Alongside the pilot, OpenAI opened two funnels for advertisers: HubSpot is its first CRM partner — customers can build campaigns, track performance, and chase leads inside HubSpot — and a ChatGPT Ads app in the Shopify App Store lets US merchants push catalogs straight into campaigns, with international availability starting September 23. There's also a natural-language Ads Manager plugin for ChatGPT Work. For regular ChatGPT users, the practical meaning: the ad layer inside the app is getting conversational.

OpenAI set DevDay 2026 for September 29 in San Francisco, announced alongside the staged GPT-6 Astra rollout. Altman's September posts had already teased DevDay as the venue where delayed launches would land, and the GPT-6 Sol rumor mill treats it as the date when any everyday-tier pricing gets confirmed. Date is set; agenda and any launch lineup remain unconfirmed.

Google is quietly piloting a program that pays publishers when their content meaningfully contributes to an AI-generated answer. The 'AI Contribution Pilot' shows up as an earnings panel inside Search Console, tied to answers in AI Overviews, AI Mode, and Gemini — the first time Google has put a dollar figure on what an AI citation is worth.

The fine print is restrictive: payouts fire only when a publisher's content 'significantly contributes' to the generation of the response — the primary logic or data shaping the answer. Being cited as a footnote, having content used only to confirm facts after generation, or fragmentary use all pay nothing. Google has approached dozens of mostly small and mid-sized publishers, and early payouts are reportedly modest — described by some participants as 'peanuts' next to traditional ad revenue.

For publishers watching search referral traffic collapse as queries resolve inside the answer box, it's a meaningful signal: influence without income is finally getting a line item, even if the checks are small and the methodology is, in one executive's words, 'quite black box.' It also sets a pricing benchmark every brand and agency can reference when arguing for GEO investment.

Anthropic is unifying Claude Chat and its Cowork agentic workspace into a single interface that auto-routes requests between them — the direct mirror of OpenAI’s ChatGPT Work and Codex merge. One conversation surface now handles both casual chat and longer agentic tasks.

Alongside it come Claude Docs and Claude Slides (beta): create and edit documents and presentations inside the conversation, with export to Google Docs, Word, PowerPoint, and PDF. Rollout starts with Pro and Max subscribers and widens over the coming weeks.

This is the consumer-app battlefield heating up in public: both labs are racing to make the chatbot the place where real documents get made. Watch adoption, not feature checklists.

Salesforce and Nvidia unveiled Koa at Dreamforce — Salesforce's first reasoning model, built on Nvidia's open-weight Nemotron and post-trained for sales, marketing, and customer-support work. The pitch: an open-weight, customer-data-free, token-cheap alternative to closed frontier models for the enterprise.