AI leaves the lab. We follow what happens next.

Archive September 2, 2026 15:08 UTC
Signal: The holes are in the plumbing, not the models - and the bill for what AI already did is landing in court.
$535 of Claude Bought a Working PLC Exploit
Forescout's Vedere Labs pointed Claude Code at a WAGO controller and got working pre-authentication code execution on live hardware - $535.74 and eight and a half hours for the final stage, and a second session permanently bricked the device.
SecurityWeek: the AI chased wrong leads until researchers moved it to Opus

Token Prices Halve, and Two IPOs Are Waiting
Silicon Data's index for what a million tokens actually trades at hit 97 cents Monday, less than half its summer high and driven by Chinese open-weight models - and OpenAI and Anthropic have both filed confidentially to go public.
Seoul Economic Daily: Kimi K3 and dynamic pricing are pulling the whole market down
NEWA .git File That Runs Code in 7 AI Agents
Manifold found eight flaws letting a received repository's own Git config run commands as the user - outside the agent's sandbox, with no approval prompt - and four were still live in Grok Build, Qwen Code, Hermes Agent and Claude Code when it retested on 1 September.
OpenAI filed three CVEs of its own the same day, covering Codex
NEW30 New Suits Say OpenAI Aided a School Shooting
Edelson PC filed 30 more complaints Wednesday over the Tumbler Ridge attack, taking the total to 37 and escalating from negligence to aiding and abetting - a claim that requires proving intent, and which OpenAI's Jason Kwon calls absolutely false.
TechCrunch: complaints name Chris Lehane as the man who told staff not to call police
Ghost Note Burning Man Won't Ban the Glasses