Edition, Fri, 11 Sep 2026.

Anthropic's threat report puts AI agents on the front line of great-power rivalry

Anthropic's most detailed threat intelligence casebook to date documents how state and criminal actors misused Claude across seven harm areas, from cyber operations to biological research and weapons development.

Anthropic's most detailed threat intelligence casebook to date documents how state and criminal actors misused Claude across seven harm areas, from cyber operations to biological research and weapons development.

Published Sept 10, the report covers activity from December 2025 through August 2026 and details disrupted operations in cyber, influence, surveillance, scams, biological misuse, conventional weapons development, and illicit model distillation. Standout cases include a Russian state-nexus actor (designated GTG-20006) that targeted more than 20 organizations in Ukraine and Europe and exfiltrated over 300,000 national identity records from a North African government authority, and suspected ShinyHunters affiliates that scanned 1.8 million Android APKs for hardcoded secrets and dumped more than 2,100 Azure AD token sets across 40+ tenants in about 34 hours.

Two Chinese undergrads used agent swarms to yield more than 12 possible zero-day findings in a month, while Alibaba's Tongyi team extracted over 151 million exchanges (across more than 3,500 fake accounts) for Qwen model training. Anthropic says it disrupted every identified operation, hardened its safeguards, and shared intelligence with authorities and partners.

OpenAI pauses its $200 ChatGPT Pro tier as Astra demand outruns capacity

For the second time in its history, OpenAI has closed a paid plan to new signups because the newest model is eating too much compute.

For the second time in its history, OpenAI has closed a paid plan to new signups because the newest model is eating too much compute.

Effective Sept 10, new sign-ups and upgrades to the Pro 20X tier at $200/month are paused, while all other plans and the API remain open and existing subscribers are unaffected. Thibault Sottiaux, who leads core products including ChatGPT and Codex, said the top tier "puts the most strain on our systems" after GPT-6 Astra, launched Sept 3, generated what he called "really unprecedented" demand.

The economics explain the squeeze: the $200 tier delivers roughly twice the Astra usage per dollar as Plus or the $100 tier, so a single Pro 20X seat can consume the peak load of 20 Plus seats, and Astra burns about 2.5x the credits of GPT-5.6 Sol. Analysts read the pause as a sign frontier capacity is still rationed, with enterprise contracts protected and consumer power users serving as the release valve.

California signs the nation's strongest rules for kids and AI chatbots

Gov.

Gov. Gavin Newsom signed a package of child-safety laws, including the first comprehensive US framework for AI companion chatbots, named for 14-year-old Adam Raine.

SB 1119 requires chatbot providers to have crisis protocols for suicidal ideation, enhanced parental controls, alerts when a child disables safety settings, and mandatory independent child safety audits with annual risk assessments. A companion law, AB 1709, bans addictive features such as autoplay and algorithmic feeds for users under 16 without parental consent.

The package also expands child sexual exploitation law to cover AI-generated material depicting anyone under 18, gives victims civil relief of up to $250,000 per action, and bans targeted advertising to children. Politico noted one of the laws is backed by OpenAI CEO Sam Altman, underscoring the political pressure now shaping AI regulation.

Positron raises $875M at a $5B valuation to bet on memory-first inference silicon

The Reno-based chip maker is skipping the industry's HBM scramble with an architecture built on commodity LPDDR5X memory.

The Reno-based chip maker is skipping the industry's HBM scramble with an architecture built on commodity LPDDR5X memory.

Positron's two-tranche Series C ($375M co-led by NEA, Atreides, Valor, Andra Capital, and Dylan Patel's SemiAnalysis Capital, plus up to $500M led by NEA and Netscape co-founder Jim Clark) lands at a $5 billion post-money valuation, with participation from Qatar Investment Authority, Cisco Investments, and Naver Ventures.

The capital funds its Asimov chip, on TSMC's N3P process with 288 GB to 2,304 GB of memory per chip and no dependence on HBM or advanced packaging, with tape-out at the end of 2026 and production in H2 2027. Its Titan system pairs 4–8 Asimov chips to serve models beyond 16 trillion parameters and 10 million-token contexts in a single node, and its first-gen Atlas system is already running 50+ racks at Oracle Cloud Infrastructure.

Cognition ships SWE-2 on an open-weight Chinese model, at a steep price cut

The Devin maker's new coding agent is built on Moonshot's 2.8-trillion-parameter Kimi K3 and costs 64% less than comparable frontier options.

The Devin maker's new coding agent is built on Moonshot's 2.8-trillion-parameter Kimi K3 and costs 64% less than comparable frontier options.

SWE-2 scores 92.8% on Terminal-Bench 2.1, 73.0% on DeepSWE 1.1, and 50.0% on FrontierCode 1.1 Main, while needing a median 18 steps per task versus 48 for Cognition's SWE-1.7. The choice to build on an open-weight Chinese foundation model is notable in a climate of US caution about Chinese AI, including CISA's distillation advisory, and frames the launch as a bet on cost and efficiency over geopolitical risk.