Anthropic AI Updates: September 11, 2026

1. Anthropic’s Threat Report Says Sophisticated Attacks No Longer Require Sophisticated Attackers

Anthropic. The September threat intelligence report covers disrupted operations across seven harm categories between December 2025 and August 2026. GTG-20006, tied to Russian espionage, compromised over 20 organizations and accessed more than 300,000 national identity records across Ukrainian and European government, defense, and drone manufacturing targets. GTG-50014 harvested 1.8 million distinct Android APKs and reached systems holding tens of millions of passenger records. GTG-10007 produced a dozen possible zero-days against security products in a single month, and GTG-50020 pivoted from hotel and fintech targets to hitting 30 AI companies in four days. The two structural findings matter more than the case list: most operations ran multi-agent frameworks with minimal human supervision, and AI API credentials are now a primary objective rather than collateral, used for “loot,” “compute,” and “cover.” Source

2. The Frontier Red Team Measured Geolocation and Drone Strike Capability, and Models Beat the Human Baseline

Anthropic. A Frontier Red Team evaluation tested intelligence targeting and weapons development across Mythos Preview, Mythos 5, Opus 5, Sonnet 5, Kimi K3, and GLM 5.2. On photo geolocation against the YFCC100M Flickr dataset, Mythos Preview hit a 37 km median error and both Mythos models beat the strongest human baseline, a Champion Division geoguessing player at 151 km. On identity correlation, models processed roughly 37,000-word samples in about 11 minutes against 2.5-plus hours for human analysts. In a simulated Betaflight quadcopter environment with 12 to 15 launches per trial, Opus 5 hit a parked vehicle 80% of the time against Sonnet 5’s 5%, and landed payloads on moving targets in wind 28% of the time while others were near zero. Anthropic’s own framing is that the gap between frontier and open-weights models “should not be mistaken for safety.” Source

3. Claude Code 2.1.268 Adds Gateway Pricing Sync and JSON Output Across Plugin Commands

Anthropic. The release wires pricing: in gateway.yaml through to signed-in Claude Code clients via managed settings, so /cost and telemetry match the spend meter. It also adds a gatewayInternalNetworks managed setting letting administrators permit /login to a Claude apps gateway on an organization’s own public IPv4 block, startup warnings when access_control.allow_cidrs is empty, claude self-hosted-runner --remove-session-state for deleting per-session directories at session end, configDirectory in claude auth status --json, --json across claude plugin install/uninstall/update/enable/disable, and Claude-chosen browser-tab icons for published artifacts. Source

4. T. Rowe Price Put Claude Cowork and Claude Code Into Its Investment Process

Anthropic. The asset manager is deploying Claude across its investment organization: Cowork for synthesizing research and running multi-step knowledge work with oversight, and Claude Code for building internal investment-support tools. The rollout sits under the AI leadership structure T. Rowe Price announced in August, with business ownership of each use case and human judgment retained on client outcomes. No adoption or productivity figures were published. Source

5. Anthropic Surveyed 1,000 Small Business Owners and Found Security, Not Skepticism, Is the Blocker

Anthropic. The Claude SMB Tour ran free half-day workshops across ten cities including Chicago, Dallas, Baltimore, and San Jose, in partnership with Tenex. A pre-launch survey of 503 decision-makers and exit surveys from 635 attendees found 81% open to new AI tools, with data security named as the top adoption barrier rather than cost or doubt about capability. Roughly two-thirds of requested use cases were about running the business rather than growing it, and nearly two-thirds of attendees asked for hands-on implementation help afterward. Eighty percent of registrants ran companies with 5 to 50 employees. Source