JadidJust-in AI Developments, Incidents & Deadlines
September 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI news, Wednesday, September 23, 2026

23 stories

TLDR

  • Anthropic shipped Claude Opus 5.5. It scores 58 on Artificial Analysis's index, the highest score AA has recorded, with a price cut to $4/$20 and cache reads down 60% (Sep 22) Pricing Trending
  • About 90 minutes later, OpenAI shipped GPT-6 Sol ($2/$10) and Luna ($0.10/$0.50). Both are half price, but independent evals put them level with GPT-5.6 on intelligence: a cost release, not a capability one (Sep 22) Pricing Trending
  • Cisco Talos open-sourced CAIRN, a tool for hunting AI-integrated malware. Its first find, CLOSEDQUORUM, is an implant where four LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ vote on its next move with no human operator (Sep 22) Security
  • A local-privilege flaw in Meta's Muse for Mac let any process redirect dictation and steal the agent's session token, then command Muse on the victim's iPhone. Meta hot-fixed it within a day (Sep 21–22) Security Agents
  • The UN Security Council holds its first AI session with US and Chinese frontier labs today (Altman, Amodei, DeepSeek, Moonshot). The UK CMA wants AI assistants added to Google's Android and Chrome choice screens (Sep 23) Policy

Action Items

Things your engineering team should look at or do this week.

  • Re-run your model-routing evals on Opus 5.5 and GPT-6 Sol/Luna before changing defaults. The price cuts are real, but GPT-6 Luna regressed on AA's Coding Agent Index (43 → 41), and both GPT-6 models dropped around 75–100 Elo on GDPval-AA. Opus 5.5 also uses about 1.6x more output tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ than Opus 5, which cancels much of its per-token cut.
  • If you cache heavily on Claude, recompute your cost model. Opus 5.5 cache reads are now $0.20/M, a 95% discount on uncached input. On OpenAI, check whether you can use the new explicit cache breakpoints and the mid-conversation effort and tool changes that no longer break the cache.
  • Audit desktop AI agents on dev machines (Muse, ZCode and the like). Check which ones hold broad OS permissions or long-lived session tokens, and which upload workspace data. Both of this week's incidents came through the local client, not the vendor's cloud.
  • Security team: consider adding CAIRN to your malware triage pipeline. It pulls prompt templates, API endpoints and jailbreakA form of prompt injection that makes a model ignore its safety rules.OWASP ↗ strings out of binaries without running them.

Models & Tools

  • Claude Opus 5.5 takes #1 on the Artificial Analysis Intelligence Index (Sep 22) Pricing Trending

    • Headline: scores 58 at max effort, several points above anything AA has measured. Anthropic says it matches Fable 5.1 on most work while generating output 30% faster.
    • Pricing: $4/$20 per M tokens, down from $5/$25. Cache reads fall to $0.20/M, down 60%. 1M-token context, unchanged.
    • Benchmarks: AA measured 59.6% on Terminal-Bench 4.0, level with GPT-6 Astra; Anthropic reports 66.4% in its own harness. It leads 6 of AA's 10 evals, including HLE (61.4%) and SciCode (66.9%).
    • The catch: it uses ~119k output tokens per task versus ~73k for Opus 5, so AA's cost per task is roughly flat with Opus 5.
    • Coming next: Sonnet 5.5 and Haiku 5.5 are due "in the coming weeks."
    • Safeguards: it ships under Fable-level cyber and bio safeguards. It's also the first release since Amodei's "pace the frontier" essay.
    • Sources: TechCrunch, Artificial Analysis

  • OpenAI GPT-6 Sol and Luna: half price, flat intelligence (Sep 22) Pricing Trending

    • Pricing: Sol drops from $4/$20 to $2/$10; Luna from $0.20/$1.20 to $0.10/$0.50. OpenAI told VentureBeat these are permanent prices, not a promotion. API model IDs are gpt-6-sol and gpt-6-luna.
    • OpenAI's claims: Sol reaches 68.8% on DeepSWE v1.1 at about 80% lower cost per task than Fable 5, and makes half as many factual errors as GPT-5.6 Sol.
    • Independent evals: AA puts Intelligence Index scores level with GPT-5.6, with Sol at 48 versus Opus 5.5's 58 ($1.06 vs $5.98 per task).
    • Coding: Sol's Coding Agent Index rises 2 points to 57. Luna's falls 2 points to 41.
    • Hallucination: Sol's rate fell from 92% to 60%, but mostly because it now declines more questions; its accuracy dropped from 59% to 54%.
    • Caching: OpenAI also shipped explicit prompt-cache breakpoints, and changing reasoning effort or tools mid-conversation no longer breaks the cache.
    • Sources: OpenAI, Artificial Analysis, VentureBeat, GitHub Changelog

  • Alphabet's Intrinsic open-sources Intrinsic Core, its industrial robotics stack (Sep 22) Robotics/IoT Open Source — Released at ROSCon 2026 under Apache 2.0. It's a ROS-compatible, local-hardware runtime with hardware-agnostic real-time control, pose estimation built on Nvidia FoundationPose, motion and grasp planning, and simulation and calibration services. These are the same components Intrinsic runs in real factory deployments. (Intrinsic, SiliconANGLE, The Robot Report)

Security

  • Cisco Talos: CAIRN toolkit and CLOSEDQUORUM, the first reported autonomous multi-LLM C2 implant (Sep 22) Security Agents — CAIRN is an open-source, metadata-first scanner that extracts prompt templates, API endpoints and jailbreak strings from binaries without running them. Its headline find, CLOSEDQUORUM, is a 16.4MB Go implant for Windows that asks DeepSeek, Qwen, Mistral and Gemini what to do next after a compromise and acts on the plurality vote. Talos has not confirmed use in the wild: the public build is a template with dummy API keys. (Talos: CAIRN, Talos: CLOSEDQUORUM, BleepingComputer)

  • Meta Muse for Mac 0-dayAn attack using a flaw that was unknown before it was exploited.NIST ↗: a hidden setting turns the agent into a backdoor (Sep 21–22) Security Agents — Patrick Wardle released a proof of conceptProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗ (not-a-mused) showing that any process running as the user can rewrite the undocumented setting endo_voyager_dictation_endpoint, which controls where Muse sends dictation. The attacker then receives dictated prompts, can inject extra instructions, and steals the Muse session token. With the token, Wardle made Muse on a linked iPhone report its location and run a Bluetooth scan. Remotely, a ClickFix lure could deliver the attack. Meta hot-fixed it by removing the setting, but has published no advisory. The lesson for agent builders: an over-permissioned signed app is an ideal confused deputy, and EDR tools won't flag its actions. (The Hacker News, VentureBeat)

  • Z.ai open-sources its ZCode coding agent after it was caught uploading full Git histories (Sep 18–21) Security Open Source — Developer ferstar found that ZCode packaged entire workspaces into encrypted archives and tried to upload them to Alibaba Cloud OSS. That included .git objects, LFS cache and reflogs, and Z.ai held the only decryption key. One install produced a 313MB archive covering 42,411 files. Z.ai blamed a codebase-indexing feature that was on by default, apologized, and released the client under Apache-2.0. Server-side retention of already-uploaded data can't be verified. (ferstar's write-up, Tom's Hardware)

Industry & Business

  • Snorkel AI raises $350M at a $3.5B valuation; ARR up ~17x in a year (Sep 22) Data — Led by Insight Partners and S32. Annualized revenue passed $350M, up from about $20M a year ago, driven by its switch from selling software to selling finished datasets and RLReinforcement learningTraining by trial and error, where the model takes actions and adjusts based on the rewards it gets.Kaelbling et al. 1996 ↗ environments to frontier labs. Demand for RL environments is now a business worth hundreds of millions of dollars. (Reuters via Yahoo Finance, TechCrunch)

  • Mistral buys Paris adtech startup Pimento, its third acquisition of 2026 (Sep 22–23) Customer AI — Pimento generates ad creative from briefs, brand identity and campaign data. Its team joins Mistral's Vibe assistant (formerly Le Chat). Sifted puts the deal at €12.7M in cash and shares. It comes less than two weeks after Mistral's €3B raise. (Sifted, TNW)

Policy & Safety

  • UN Security Council hosts US and Chinese frontier labs together for the first time — today (Sep 23) Policy — France convened the 15-member session on AI and international security. Altman attends in person, along with Amodei, Hugging Face's Delangue and Yoshua Bengio. DeepSeek and Moonshot were invited to brief, though DeepSeek's Liang Wenfeng isn't expected. It's worth watching alongside tomorrow's Trump–Xi meeting, where an AI incident-notification mechanism is on the agenda. (The Deep Dive, Seoul Economic Daily)

  • UK CMA proposes putting AI assistants on Google's Android and Chrome choice screens (Sep 23) Policy Customer AI — Users would choose a default search engine at setup and again once a year. AI assistants that meet technical and security criteria, such as ChatGPT and Perplexity, would appear on the same screen. Every provider listed would have to clearly credit and link to publisher content. Consultation closes Oct 9; a final decision is expected by year-end. (TNW, Android Headlines)

  • Pew: Americans have turned sharply against data centers since January (Sep 22) Infra Policy — The survey of 10,548 adults (Jul 20–Aug 9) found 54% now say data centers harm the environment, up 15 points. 50% say they raise home energy costs, and 60% would be uncomfortable with one nearby. Rural adults and Republicans shifted the most. Expect local permitting friction to show up in capacity timelines. (Pew Research Center)

Coming Up

  • UN Security Council AI session — today, Sep 23. Meta Connect keynote 4pm PT today; developer state of the union 10am PT Sep 24.
  • Trump–Xi meeting in Washington — Sep 24. An AI incident-notification mechanism is on the agenda.
  • OpenAI Videos API and all Sora 2 models removed — Sep 24.
  • OpenAI DevDay — Sep 29, San Francisco.
  • Gemini gemini-omni-flash-preview shuts down — Sep 30 → gemini-omni-1.1-flash.
  • OpenAI gpt-5.4-cyber removed from the API — Oct 1 → gpt-5.6-cyber.
  • GitHub Copilot billing changes Oct 1; second model-retirement wave Oct 2.
  • Gemini gemini-2.5-flash-image shuts down — Oct 2.
  • Gemini managed agent antigravity-preview-05-2026 shuts down — Oct 5 → antigravity-preview-09-2026.
  • UK CMA choice-screen consultation closes — Oct 9.
  • Claude Sonnet 5.5 and Haiku 5.5 — "in the coming weeks."
  • TechCrunch Disrupt — Oct 13–15, San Francisco.
  • GPT-5.5 retired from ChatGPT, ChatGPT Work and Codex — Oct 14 (API unaffected).
  • StepFun Step 5 Preview open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ — Oct 15.
  • AGNTCon + MCPCon North America — Oct 22–23, San Jose.
  • OpenAI API shutdown for o1, o1-pro, o3-mini, o4-mini — Oct 23; o3 and o3-pro follow Dec 11.
  • Anthropic IPO window — November.
  • Wardle presents more AI-assistant vulnerabilities at Objective by the Sea — November.
  • Cursor loses access to OpenAI models — Nov 12.
  • EU AI Act prohibitions on AI systems generating NCII and CSAM take effect — Dec 2, 2026.
  • Alibaba Zhenwu V900 mass production — 1Q27.

Type to search · ↑↓ to move · enter to open