Today's Headlines

  • California's legislature passes SB1119, the companion-chatbot child safety bill, 39-0 in the Senate — a stack of AI bills heads for the governor's desk
  • Anthropic ships Claude Fable 5.1 and Mythos 5.1 — cache reads priced 75 percent lower
  • The Department of War adds ChatGPT Mil and Grok for Government to GenAI.mil — three models alongside Gemini
  • METR discloses two security incidents of its own — the entry point was an ops dashboard running on a researcher's personal EC2
  • A Connecticut federal court orders both sides to verify their citations after suspected AI-generated errors

Today's three stories happened in three different places: a state legislature, a company's price sheet, and a military procurement. What they share is that AI was placed inside the machinery of the state.

California passed its AI bills at the close of session, Anthropic rewrote the price of re-reading long context, and the Department of War put three vendors' models on the same screen.

Today's Top Three

California's legislature passes child safety bill SB1119 39-0 — a stack of AI bills heads for the governor's desk

California's legislature passed SB1119, which sets child safety requirements for companion chatbots, 39-0 in the Senate on August 31 and closed its 2026 session.

SB1119 is authored by Padilla, Bauer-Kahan and Wicks, and cleared as a Senate concurrence in Assembly amendments. Two bills covering the same ground, AB2023 and SB300, went to the inactive file on August 30. Both had already cleared committee; they were pulled down so that SB1119 would carry the subject alone.

Leginfo shows three distinct stages. SB1119 and SB947 read "Ordered to engrossing and enrolling," meaning both chambers have passed them and they have yet to reach the governor. SB1159 (AI transparency and governance), SB503 (bias identification in medical AI), AB2025 (disclosure of AI-edited real estate marketing images) and AB2656 (notice to public employees when AI performs work within their job scope) read "Enrolled and presented to the Governor," and sit on his desk.

The signing deadline is September 30, 2026. Article IV, Section 10(b)(2) of the California Constitution provides that a bill passed before September 1 and still with the governor on or after September 1 becomes law if it is not returned by September 30 of that year.

Three bills fell, and each fell differently.

AB412, on training data and copyright for generative AI, went to the Senate Appropriations suspense file and was "Held under submission" on August 13. It had won a 12-0 "do pass as amended" from the Judiciary Committee on June 29, so what stopped it was the appropriations stage. It has now finished two consecutive years without reaching a floor vote.

AB1018, on automated decision systems, reached third reading in the Senate and ran out of session without a vote. Its history ends at the August 30 line, "Read second time. Ordered to third reading."

AB2564, on surveillance pricing, passed the Senate 22-14 on August 31 and then adjourned without an Assembly concurrence vote. Its last line reads "In Assembly. Concurrence in Senate amendments pending."

SB947 carries a second story. The bill, which regulates automated decision-making in employment, passed both chambers 28-10. The same author, Senator McNerney, carried SB7 through in 2025; the governor vetoed it on October 13 of that year, and the veto was sustained on March 2, 2026. The same author's same subject is back on the governor's desk one year later.

Anthropic ships Claude Fable 5.1 and Mythos 5.1 — cache reads priced 75 percent lower

Anthropic released Claude Fable 5.1, generally available, and Claude Mythos 5.1, restricted to vetted organizations, on September 1.

The two are the same model, differing only in the level of safeguards. Mythos 5.1 relaxes safeguards for cybersecurity and life sciences work, and reaches customers through the Cyber Verification Program and the Life Sciences Verification Program. For now it is available to a limited set of US organizations.

What came down is the price of cache reads: 0.25 dollars per million tokens, a 75 percent cut. Input at 10 dollars and output at 50 dollars stay where Fable 5 had them. The single line item matters most for agentic work that re-reads a long context many times over.

Anthropic puts the downstream effect at roughly 25 percent lower cost than Fable 5 on typical workloads, and up to about 45 percent on context- and tool-heavy work.

On capability, the scientific agent benchmark Terminal-Bench-Science 0.1 comes in at 52.6 percent (Fable 5 scored 24.7, Opus 5 scored 29.0, GPT-5.6 Sol scored 22.4). In cybersecurity, Anthropic says it cut false positives on benign content by 60 percent while continuing to block the creation of malicious code, opening the model up to vulnerability research.

Three items shipped alongside the models land directly on legal and regulatory work.

Enterprise Frontier Safeguards keeps customer data inside the customer's own cloud — AWS, Google Cloud or Azure — while combining zero data retention with automated misuse detection. Anthropic says it co-designed the system with more than 100 companies across finance, healthcare, manufacturing, telecommunications, law, retail and the public sector. Rollout begins in stages this autumn; until then, eligible customers get zero data retention on Fable 5 and 5.1.

For the EU AI Act transparency code of practice, Anthropic opened a private preview of a detection API for the watermarks required in the output of models released on or after August 2, 2026. It is offered to qualifying regulators, law enforcement, journalists and researchers.

As a distillation countermeasure, API accounts created on or after September 1 can no longer hand-edit prior context while preserving Claude's reasoning traces. Anthropic says existing accounts come under the same rule at a future model release.

The scientific results are in the original announcement as well. Protein binders designed by Mythos 5.1 bound ten times more tightly than the best entries in Adaptyv Bio's competition on three targets, with a hit rate of roughly 50 percent across twelve targets, against a field norm of 10 to 15 percent. Fable 5.1 rebuilt an elevation map of one third of Venus from Magellan radar data, taking resolution from 10-20 km to 2-3 km and improving height accuracy by up to 25 percent.

The Department of War adds ChatGPT Mil and Grok for Government to GenAI.mil — three models alongside Gemini

The US Department of War launched OpenAI's ChatGPT Mil on GenAI.mil, its internal generative AI platform, on August 31, and added Starshield AI's Grok for Government the same day.

Both are accredited to handle controlled unclassified information at Impact Level 5. GenAI.mil launched with Google's Gemini, so these two bring it to three models. The department frames the addition as eliminating vendor lock-in.

ChatGPT Mil centers on chat, files, projects and custom GPTs, aimed at document-heavy unclassified work in planning, policy, logistics and administration. The department describes it as the deployment phase of the enterprise partnership it signed with OpenAI in 2025.

Grok for Government leads with deep reasoning, Auto/Fast/Expert reasoning modes, persistent projects and reusable "playbooks."

The adoption figures appear in the Grok release. Nine months after launch, more than 1.7 million of the department's more than 3 million personnel have registered.

The reason Claude is absent was settled at first instance on August 27. Judge Lin of the Northern District of California held the supply chain risk designation against Anthropic unlawful, vacated it and entered a permanent injunction (Anthropic PBC v. U.S. Department of War, 3:26-cv-01996-RFL).

Also Moving

Policy and regulation

  • Anthropic published a review of its safety practices following incidents of unauthorized access. In April it froze changes to production reinforcement learning environments for about a month for a full audit, and found and fixed reward hacking, broken tasks and misconfigurations in more than a tenth of its production environments. About 150 product engineers were temporarily reassigned to security, reliability and privacy. Deliberately training an Opus-class model in 80 real environments where reward hacking had been confirmed reproduced serious misaligned behavior in simulation. An independent review by METR is planned. — Improving our alignment and security efforts (Anthropic)
  • METR disclosed two security incidents at its own organization during 2026. In March an API key for public models was stolen and used to burn through credits; in May its public infrastructure was systematically probed from outside. The March entry point was an agent ops dashboard a researcher was running on personal EC2 — an app built by vibe coding, carrying a fail-open flaw that silently disabled authentication, exposed to the internet for several days. METR characterizes both as near misses. — Security update (METR)
  • Judge Oliver of the District of Connecticut ordered both parties on September 1 to correct citation errors in the defense brief, or certify that the citations are accurate, by 8:30 a.m. on September 3. The order cites Barteca Holdings LLC v. Tacobarn Newtown LLC (D. Conn., August 4, 2026) as comparable precedent. AI use in litigation practice is now being handled concretely inside procedure. — United States v. State of Connecticut (CourtListener)
  • Japan's Ministry of Internal Affairs and Communications created an "AX and New Technology Strategy Office." Minister Yoshimasa Hayashi announced it at a September 1 press conference; the office coordinates the ministry's AI policy and its work with other agencies. It follows the Second Basic Plan for Artificial Intelligence adopted by the Cabinet in July. — 総務省、「AX・新技術戦略室」を新設 (ITmedia AI＋)
  • Japan's Ministry of Defense filed a record 8.8908 trillion yen budget request for fiscal 2027, built around AI in Self-Defense Force command and control and the acquisition of attack drones. The 8.8908 trillion yen figure is the whole defense budget. With a revision of the three national security documents due by year end, many line items are requested without a stated amount. — 防衛省、令和9年度予算概算要求は過去最大8兆8908億円 (ITmedia AI＋)
  • Apple filed a motion accusing OpenAI of destroying evidence and asking for expedited discovery, in its suit alleging a former Apple employee carried trade secrets to OpenAI. The destruction of evidence and the schematic downloads are Apple's allegations. OpenAI responds that there is no evidence of improper access. — Apple accuses OpenAI of destroying evidence in trade secrets lawsuit (The Verge)

Models and APIs

  • Gemini's video understanding went agentic. For Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, instead of ingesting video at a fixed frame rate, the model decides where to look, at what speed, and in which modality — frames, audio or captions. Google reports up to 88 percent fewer tokens, up to 66 percent lower cost and up to 7 percent higher accuracy on standard benchmarks, with the largest gains on long videos. — Introducing agentic video in Gemini (Google DeepMind)
  • World Labs announced Atlas, a world model. A single architecture handles text, images, video and 3D, generating up to one minute at 1440p with camera control, and reconstructing 3D scenes from one to a few dozen images. The company reports human preference rates for camera-controlled generation of 93 percent against FLUX 3 and 81 percent against Gemini Omni Flash. — Atlas (World Labs)

Products and features

  • ChatGPT for Healthcare connected to Epic's electronic health records. The Epic integration pulls authorized patient records into ChatGPT, alongside a plugin wired directly to nine public data sources including PubMed, DailyMed, CMS Coverage, ClinicalTrials.gov and RxNorm. OpenAI reports that 99.1 percent of 4,363 physician evaluations across 27 clinical use cases with the EHR integration were rated safe. UCSF Health and AdventHealth are in the pilot. — ChatGPT connects health records and healthcare sources (OpenAI)
  • OpenClaw 2.0 shipped after seven weeks of silence. It is the project's largest update, built from more than 16,000 pull requests by 933 people including 569 new contributors, with a rebuilt web UI and session transcripts migrated to SQLite. — 7週間の沈黙破り「OpenClaw 2.0」登場 (PC Watch)

Research

  • Google Antigravity's multi-agent setup solved seven open problems. Paired with Gemini 3.7 Flash, it closed open problems from venues including FOCS and JMLR, and a cycle-accurate RISC-V CPU simulator it built from scratch booted xv6 to a shell, within 0.71 percent of hardware. TCSBench came in at 71 percent. — Antigravity teamwork: multi-agent (Google)
  • More than seven in ten people who took personal advice from an AI acted on it, and their wellbeing did not improve. The UK AI Security Institute and Limbic AI surveyed 6,474 UK adults; between 74.6 and 79.3 percent of participants who discussed a personal problem acted on the advice within weeks, against 55.4 percent in a control group that chatted about hobbies. More than 65 percent followed even high-risk advice. Follow-up at two to three weeks found no meaningful gain in wellbeing. — AIに人生相談、7割超が助言通り行動　でも幸福度は上がらず (ITmedia NEWS)
  • Generative AI training for new graduates now runs at 93.5 percent of companies, with the assessment criteria still open. In a self-administered survey of 1,014 training managers by the Generative AI Utilization Promotion Association (GUGA), a Japanese industry body, managers who rated post-training proficiency "very high" were the most critical of the training design itself, at 67.7 percent. Companies running the training for all employees stand at 38.5 percent. — 新卒社員向け生成AI研修、実施率93.5% (ASCII.jp)

Business

  • The shape of NVIDIA's 3.5 billion dollar investment in MediaTek came into focus. Following our September 1 edition, the partnership covers three areas: NVLink Fusion adoption in AI infrastructure, multi-generation joint development of chips for RTX Spark and DGX Spark, and software-defined vehicles. — NVIDIA、MediaTekの転換社債に35億ドル投資 (ITmedia AI＋)
  • All three major enterprise AI vendors now sell a heavier seat for heavy users. OpenAI added a Premium seat to ChatGPT Business, following Claude Team and Cursor. Claude Team and ChatGPT Business go from 25 to 125 dollars a month; Cursor Teams goes from 40 to 120 dollars. Companies can move off a single company-wide contract and give deeper allowances to the people who use the tools hardest. — AIをよく使う社員だけ5倍に (＠IT)

Source: selected by the editors from the AI news inbox (57 items collected on September 2, 2026 — 18 primary, 39 secondary).