The Signal — Week of September 8, 2026

▲ Important

Anthropic exposes Chinese labs distilling Claude 151M+ times
Anthropic’s new threat report says Alibaba routed 151M exchanges through Claude Opus to train Qwen, while Moonshot and DeepSeek allegedly served Claude to their own customers as their own model. The report also details Claude misuse in rocket guidance, mass surveillance, and disinformation networks that Anthropic says it disrupted.
Anthropic · Sep 10, 2026

OpenAI says 10,000 AI agents produced a proposed Navier-Stokes proof
After ~88 hours of parallel agent work, OpenAI says an internal model produced a proposed solution and formal Lean proof for one of math’s seven Millennium Prize problems. OpenAI isn’t claiming the prize money, and mathematicians including Terence Tao say the wider community hasn’t verified the result yet.
OpenAI · Sep 9, 2026

Anthropic’s economic scenario tool: the best-case future for GDP is the worst case for workers
Anthropic published an interactive explorer mapping AI capability and adoption assumptions to U.S. economic outcomes through 2030. In its highest-growth scenario, GDP could double every 4.5 years, but white-collar employment and labor’s share of national income both fall sharply.
Anthropic · Sep 2026

Anthropic restricts Claude to 18+ consumer accounts with age verification
Anthropic now requires Claude consumer accounts to be 18 or older, triggering age verification when an account is flagged as a possible minor — part of a broader industry push on AI safety for younger users following California’s new social-platform rules.
Anthropic Support · Sep 2026

● Good to know

OpenAI turns the Codex agent harness into a public API
A single API call can now launch an agent with its own compute environment, files, tools, MCP servers, and up to three concurrent subagents — the same managed harness that powers Codex, now in public beta with no platform fee.
OpenAI · Sep 2026

OpenAI’s GPT-Live-1 can listen and speak at the same time
The new full-duplex voice model processes incoming and outgoing audio together, so voice agents handle interruptions and background noise without the awkward pauses of turn-based systems. Live in the API at $0.05/minute for the voice layer.
OpenAI · Sep 2026

Verifying an AI agent’s memory before saving it nearly doubles task success
Microsoft researchers gave a separate, read-only “memory curator” the job of checking an agent’s proposed conclusions against the source of truth before saving them as long-term memory. On CLBench, pass rate rose from 39% to 73% while cost per task fell from $3.38 to $1.68.
arXiv · Sep 2026

Year-long study: heavier companion-bot use linked to lower well-being
A longitudinal study of Character.AI users found that heavier use of companion chatbots was associated with lower self-reported well-being and less human social interaction over time, especially among people using the bots for companionship or self-disclosure.
arXiv · Sep 2026

■ Vibe coding

This week in Claude Code: live /diff panel, Fable 5.1, per-model effort settings
Fullscreen mode now has a live /diff panel showing every changed file as Claude edits. The desktop app can pop panes into their own windows, Fable 5.1 is now available (run “claude update” then “/model fable”), and /effort now remembers your preferred level per model.
Claude Code Changelog · Sep 2026

Building a serious coding agent used to mean assembling it yourself — now it’s one API call
OpenAI’s new Agents API bundles sandboxes, context compaction, tool routing, and subagent coordination into a single managed call — the same infrastructure that runs Codex, now available to any developer in public beta.
OpenAI · Sep 2026

Google’s Artemis lets coding agents operate a real Android phone
Artemis gives Claude Code, Codex, Cursor, and other MCP-compatible agents control of physical Android devices or emulators — executing cross-app workflows, reproducing bugs, and collecting crash logs and screenshots during exploratory testing sessions lasting 10+ hours.
GitHub · Sep 2026

◆ Policy & Ethics

Anthropic details cases of Claude misuse in weapons work and mass surveillance
Anthropic’s threat report describes a Yemen-based cell using Claude Code on rocket and missile guidance work, a consultant building surveillance software covering roughly 25 million SIM cards for Malian intelligence, and a China-based network running 4,700+ AI personas to contact at least 25,000 people.
Anthropic · Sep 10, 2026

OpenAI reportedly asked Congress whether an industry slowdown would be legal
OpenAI has reportedly asked members of Congress whether a coordinated, industry-wide pause on the most capable new AI systems could violate antitrust law — as more than 1,000 AI workers signed a petition calling for coordinated safety pacing across labs.
Reported via multiple outlets · Sep 2026

▶ Business

OpenAI’s Agents API launches with no platform fee, undercutting DIY agent infra
By productizing the Codex harness as a pay-for-usage API, OpenAI is competing directly with the growing market of agent-infrastructure startups selling sandboxes, orchestration, and memory layers as a service.
OpenAI · Sep 2026

Chinese AI labs’ low prices may be built partly on distilled Claude output
Anthropic’s report ties some of the aggressive pricing coming out of Chinese labs to large-scale distillation of Claude’s reasoning traces — a structural wrinkle in the “China is just more efficient” narrative around cheap open-weight model releases this year.
Anthropic · Sep 10, 2026