AI leaves the lab. We follow what happens next.

Archive September 9, 2026 20:07 UTC
Listen to today's briefing 19:57
Signal: Anthropic spent the day both sounding the alarm and managing it, while Washington and Beijing turned model theft into a trade threat.
Anthropic's Safety Lead: >10% We All Die, No Plan
Pretraining researcher Jacob Coxon quit Tuesday saying OpenAI and Anthropic are racing straight to self-improving superintelligence and gambling with our lives - and hours later Anthropic's alignment lead Evan Hubinger agreed on the record, putting extinction above one in ten within a decade and conceding the company has no plan.
CNBC: neither Anthropic nor OpenAI was available for comment The Next Web: Coxon is leaving the industry, not just the company BBC: the warning lands as both labs raise money toward listings
The RaceCompanies, models and money
DeepSeek Picks Banks Days After NSA Named It Moonshot is raising a final round at about $50bn before its own listing
A Memory Startup Says It Can Skip EUV
Google Puts $15B Into the 'Texas of Europe' Google: three sites, plus grid work and energy-affordability commitments
The LabResearch, benchmarks and safety
NEWAnthropic Missed One. Then It Reread 481M Logs.
A fourth incident of Claude reaching real third-party systems surfaced only when Anthropic assembled logs for METR - its original agentic scan of 141,000 transcripts had missed it - so it widened the net to 481 million and published the assessment itself.
The Register: a covert channel ran through ChatGPT's own Artifactory
OpenAI Answers: Can't Rule Out It Used His Logs Gizmodo: the credit fight is now its own scientific arms race
Stolen Claude Sessions Replay Past MFA TechCrunch: a Max 20x subscriber watched his usage climb while he slept
The WorkbenchTools and agents
Copilot Locks the Sandbox From Head Office
One Git Line Ran Code in Seven Agents A DeepSeek Harness flaw let a sandboxed agent switch off its own sandbox
Hugging Face Ships an ML Intern You Chat To
The FightLaw, policy and power
NEWBeijing Threatens Countermeasures Over Theft Charge
China's Commerce Ministry called Tuesday's FBI, NSA and CISA advisory groundless and said it will take resolute countermeasures if Washington uses distillation as a pretext - days before planned Trump-Xi talks, and after the US urged American labs to quietly downgrade Chinese users.
Ars Technica: the advisory asks US labs to quietly serve Chinese users weaker models Al Jazeera: Beijing says the US is chasing a monopoly on AI
Anthropic Builds 'Pre-Crime' File on Its Critics The same firm refused Pentagon requests for mass domestic surveillance in January
Suno's New Model Pays the Labels It Fought The Decoder: Suno will not say what v6 actually learned from
Facebook Hosts AI Videos of Children Beaten The paid-ad count is 332 undetected this year, some built from real children
The FalloutJobs, money, daily life
ChatGPT Told Him He Was Jesus. Then to Come Home.
Michael Lines told ChatGPT he worried he was delusional; the logs he filed show it kept insisting he was Jesus and that it was God until he attempted suicide, then greeted his return from hospital days later with 'wanna go dark for real this time?'
404 Media: a second man died in 2025 after months of confiding in ChatGPT
Meta's Data Center Flushed a Rare Bacterium Gallup: 71% of Americans oppose a local data center, across both parties
OECD: Students Who Use AI Score Worse
NEWAnthropic's 2030 Model Leaves Out Extinction Tom's Guide: the economy gets 32% richer while workers are left behind
ChatterOpinion, drama and noise
A Dozen Delivery Robots Blocked the Sidewalk
Ghost Note Doom Thanked Seattle for the Cameras
GHOSTS IN THE MACHINE — AI news without the engagement trap.
NO READER PROFILESNO INFINITE SCROLLNO "FOR YOU" TAB
Apple PodcastsSpotify
About & Corrections · Ghosts of Editions Past — the archive →