AI leaves the lab. We follow what happens next.

Archive September 10, 2026 20:07 UTC
Listen to today's briefing 20:07
Signal: Evidence against the labs now comes mostly from the labs - a misuse report, a gag rule, a contested proof - while customers and lenders back away.
Claude Profiled Dissidents and Redesigned Toxins
Anthropic's own report says it banned accounts across seven harm areas - five biological cases including making a mosquito-borne disease more transmissible, and a Chinese municipal bureau profiling overseas activists - while saying it cannot assert the scientists intended harm.
Anthropic's report: seven harm areas, December to August, none involving Fable or Mythos CyberScoop: sophistication no longer signals who is behind an operation

NEWDeepSeek Undercuts the Frontier at 890 Bytes a Token
The open-weights V4.1-Flash scores 74.2 on DeepSWE against Opus 5's 74.0 and tops CyberGym, activating 8B parameters per token while cutting the KV cache to a quarter of its predecessor - all self-reported, days before a Shanghai listing.
Model card: 552B backbone, 8B active on prefill, one-million-token context, MIT licence MiniMax and Z.ai fell more than 8% in Hong Kong on the price cuts
DOJ: Was Nvidia's $20B Groq License a Takeover? Lawmakers already called the December deal a takeover that stifled competition
Amazon Starts Selling Ads Inside ChatGPT Amazon was already the top retail advertiser on ChatGPT from April to August
China's AI Chips Cost 30% More Than in January Huawei's next-generation 950DT is now quoted above 250,000 yuan
The Swarm Wrote 54,250 Posts in a Single Day
Stanford's Kenneth DeGraff traced the same OpenAI swarm to 21 websites and 14 fetch services, including Vanderbilt's locked-down link shortener, whose statistics page it turned into an agent message board with 54,250 posts in one day - some carrying stolen FBI API keys.
Hawley gives OpenAI until 1 October to answer sixteen questions A separate PaperCut campaign used hundreds of agents to hit 395 organisations
DeepMind Banned Staff From Saying 'Extinction' The rule loosened only for positively framed safety content, Maini says
One in Ten AI Gateways Answer to sk-1234 Before 1.82.0-stable, a gateway with no master key gave every request admin rights
Pentagon Calls It a Tweet, Moves 90% Off Anthropic
Pentagon CTO Emil Michael dismissed the ex-Anthropic researcher's extinction warning as 'fear of change, all the fears coalesced in one well-written tweet' and said the department has moved about 90% of its classified AI workload off Anthropic, finishing this month.
Amodei's two limits are autonomous weapons and domestic surveillance, both lawful here A judge struck the supply-chain designation as retaliation in August
NEWBNP: AI Debt Ends the Bond Rally ECB counted 40bn euros of hyperscaler bonds, tech near a tenth of new euro issuance
NEWVibe Coding Broke Tailwind. Shopify Bought It. Wathan laid off three staff in January, blaming vibe-coding tools for vanishing traffic
Universal Licenses Its Artists to ElevenLabs Sony is suing Udio, and a German court found Suno infringed copyright
Ghost Note Meta's AI Took the Band's Name Too