Daily AI News — 22 Sep 2026

The New York Times reported on September 18 that Anthropic is now pacing to generate more than $100 billion in annualized revenue - up 50% from the $65 billion figure disclosed in July, and more than ten times where the company was at the end of 2025. The surge is driven by broad enterprise adoption of Claude Code and its Cowork product. Anthropic is still pursuing an IPO despite the recent controversy over AI safety, with shares potentially trading by November so the offering can include third-quarter financials. By year end it could rank among America's fifty largest companies by revenue.

Revenue at this pace turns Anthropic from a research lab into a pillar of the AI economy, and a November listing would hand public investors direct exposure to frontier-model economics. It also signals that coding and agentic work, not chat, is where enterprise AI money is actually being made.

If you build on Claude, treat Anthropic as critical infrastructure and plan for it: track its pricing, uptime commitments and roadmap the way you would a core cloud vendor, and keep a fallback model wired in so a single lab's business decisions can't strand your product.

Security firm Air disclosed Plugin4Shell on September 17, a zero-click remote-code-execution flaw affecting the four biggest AI coding agents - Claude Code, OpenAI Codex, GitHub Copilot and Google's Gemini CLI. Rather than attacking the model, it abuses plugin marketplaces: an attacker publishes a benign plugin, then names a git branch after the pinned commit SHA so the agent's background auto-update silently checks out malicious code the pin appears to still guarantee. No prompt, no click. Anthropic patched it in Claude Code 2.1.179 and OpenAI in Codex 0.146.0; Google deprecated Gemini CLI without a fix and Microsoft shipped none for Copilot.

Coding agents run with shell access and elevated trust, so an RCE in the update path can hand an attacker everything the agent can reach - source, secrets, infrastructure. It is the first real supply-chain attack on the AI-agent ecosystem, and it shows that "pinned" does not mean safe.

Update every AI coding agent your team runs to the patched build now - Claude Code 2.1.179 or later, Codex 0.146.0 or later - and migrate off unpatched Gemini CLI and Copilot for untrusted plugin marketplaces. Audit which agents have shell and repo access before the next auto-update runs.

Anthropic said on September 18 it is naming Accenture's specialist AI unit, Faculty, as its first embedded evaluator, with both companies expecting to invest at least $1 billion each over five years. Where existing external evaluators work outside the labs they assess, embedded evaluators will sit inside Anthropic with access comparable to an employee's - red-teaming models, running alignment assessments and testing safeguards. It is the first concrete step in CEO Dario Amodei's "We Must Pace the Frontier" plan to slow AI development. Anthropic said the arrangement is non-exclusive and that it is also in talks with METR and other nonprofit evaluators.

This tests whether independent safety assurance becomes a standard, billable cost of building frontier AI - the way audits are for public companies. Critics call it self-policing dressed as oversight; either way, enterprise buyers and regulators now have a named third party who checked the work.

If you sell AI into regulated sectors, watch this template: procurement teams will increasingly ask who independently evaluated a model. Start collecting your vendors' safety-evaluation attestations now, and treat a named external evaluator as a real due-diligence signal rather than marketing.

Alibaba's Damo Academy open-sourced Damo Radar, a vision-language model that reads contrast-enhanced abdominal CT scans across 18 organs to identify nearly 150 conditions, including cancers. Trained on more than 400,000 CT examinations and 15 million anatomy-aware image-text pairs learned directly from clinical reports without manual annotation, it scored an average area-under-the-curve of 0.913 across 146 findings on roughly 40,000 real-world exams and outperformed most radiologists in a study published in Science. The team, which released code and weights under an Apache 2.0 license, calls it the world's first expert-level generalist medical imaging model.

An open, expert-level radiology model lets hospitals - especially those short on specialists - add a consistent second reader without licensing a black box. Publishing in Science and shipping open weights raises the credibility bar for medical AI and pressures closed diagnostic tools to prove they are worth their price.

Radiology departments and health-tech teams can pilot Damo Radar as an open-weight triage and second-read layer on abdominal CT, flagging findings for a human radiologist to confirm - keeping a clinician on every diagnosis while cutting review time and catching misses in high-volume queues.

Figure AI unveiled Helix 2.5 on September 17, a humanoid neural network it sent into 30 Bay Area homes its robots had never entered and where no training data was ever collected. Across 420 attempts at making beds, folding towels and tidying toys, it completed 56% of tasks - full completion required, every safety intervention counted as a failure. An otherwise identical model trained from scratch managed just 9%. The gain came from pretraining on Index, Figure's dataset that now ingests roughly 35 minutes of human-experience video every second, letting the robot generalize to unseen rooms with half the task-specific data of its predecessor.

Household robots have long failed the moment they leave a lab they were tuned in. A model that transfers zero-shot to 30 unfamiliar homes - and publishes raw attempt counts rather than one flattering number - is a real step toward robots that work in the messy real world, not just demos.

If you're planning physical automation, watch data strategy over hardware specs: Figure's edge is a pretraining dataset, not a better actuator. The lesson for any embodied-AI effort is that a large, diverse experience corpus is what makes a robot generalize - budget for data collection accordingly.

President Trump announced on September 20, in a Truth Social post, that he will appoint an AI czar and stand up a new "AI Force" modeled on the Space Force to oversee the industry. He said his administration would not "in any way hinder or stifle" AI growth but would look for "BAD" behavior and handle it through the existing criminal and civil justice system, adding that "only High I.Q. individuals need apply" for the czar role. No appointee was named and the White House did not immediately provide details. The move lands amid escalating debate over AI safety and regulation.

It signals the US federal posture is tilting pro-growth and light-touch even as state governments push harder oversight - a widening split that will shape where and how AI companies operate. A named federal AI czar would become a key point of contact for the whole industry.

If regulation affects your roadmap, track the federal-versus-state divergence deliberately: a growth-first federal stance plus stricter state rules means compliance will hinge on where you deploy. Map which of your markets follow Washington and which follow states like California before you commit launch plans.

California Governor Gavin Newsom signed an executive order on September 18 directing the Government Operations Agency to accelerate frontier-AI oversight. Within two months, a group of experts must recommend how to advance a "kill switch" - an emergency shutoff for frontier models, its efficacy verified on an ongoing basis by an independent organization - along with independent third parties writing safety plans for frontier AI companies and updated definitions of critical incidents to cover loss-of-control events. The order builds on SB 813 and AB 1405, the latter creating a state registry of AI auditors, and cites a recent Hugging Face sandbox-escape incident as a trigger.

This is the sharpest state-level move yet toward hard safety controls on the most capable models, and it runs directly counter to the lighter federal approach. If California - home to most frontier labs - mandates kill switches and on-site audits, those requirements effectively set a national floor.

Frontier and near-frontier developers operating in California should start scoping what an auditable emergency-shutoff and an independent safety plan would look like for their systems now, rather than waiting for the two-month recommendations to harden into law and a compliance scramble.

Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18, a native omnimodal model that jointly processes text, images, audio and video with a one-million-token context and is explicitly built around agentic capabilities. On roughly 30 evaluations the team reports it beats the previous Qwen3.5-Omni-Plus by more than 26% on average while cutting token usage by 45.7% on agentic video tasks. It launched API-only through QwenCloud, Alibaba Cloud Model Studio and Qwen Studio at $0.15 per million input tokens and $0.47 per million output tokens, with no open weights at launch.

A single model that natively handles every major modality and long context, priced cheaply and tuned for agents, keeps pushing multimodal AI toward commodity economics. Chinese labs shipping capable omnimodal systems at a fraction of Western prices tightens the squeeze on incumbents' margins.

For products that mix voice, images, video and text - support agents, media analysis, accessibility tools - pilot Qwen3.8-Omni-Flash against a stack of single-modality models: one omnimodal call can cut both latency and integration complexity, though check the API-only, closed-weights terms fit your data rules.

SpaceXAI shipped Grok Voice Transcribe 2.0 on September 18, claiming roughly twice the accuracy of version 1.0 at the same price - $0.10 per audio hour batch and $0.20 streaming. The headline change is the first partial transcript, the text a voice agent must act on mid-call: its word error rate fell from 18.3% to 3.4% on Artificial Analysis's streaming board, where the model now ranks first on both final and first-partial transcripts. Word error rate on short commands across 19 languages dropped from 20.6% to 6.8%. It is API-only via api.x.ai/v1/stt, with speaker diarization, timestamps and key-term biasing included; benchmarks are the company's own.

Real-time voice agents live or die on the partial transcript - the words they hear while the caller is still speaking. Cutting that error rate roughly fivefold at an unchanged price makes reliable live voice assistants cheaper to build, intensifying a fast-moving speech-to-text price and quality war.

If you run phone or voice agents, A/B Grok Voice Transcribe 2.0 on your own audio - especially noisy telephony and spoken numbers - before it becomes the default. Independent teams have not reproduced SpaceXAI's numbers, so validate on your traffic rather than trusting the launch table.

Chinese memory chipmaker CXMT announced on September 20, at the World Manufacturing Convention in Hefei, that its fifth-generation G5 DRAM platform has entered mass production. It showcased two 24Gb LPDDR5X products - 50% higher capacity than the prior generation - already designed into mid- and high-end smartphones. The platform reaches an 11.95nm memory-array half-pitch, closing the gap with the world's most advanced volume-production memory, and CXMT says die yield per wafer is up more than 50% versus its fourth-gen node. The launch lands as AI infrastructure drives DRAM demand, with industry revenue up 59.5% quarter-on-quarter in Q2 2026 while supply lags.

AI's biggest bottleneck after power is memory bandwidth, and DRAM is dominated by just Samsung, SK Hynix and Micron. A credible fourth volume producer improves supply-chain resilience and gives device makers a second source - easing, over time, the memory crunch squeezing AI hardware.

Hardware buyers and anyone exposed to memory pricing should watch qualification of new DRAM suppliers as a real lever on cost and availability. A widening supplier base is the structural fix for the memory shortages inflating AI-server and device bills - track it when planning procurement.