AI leaves the lab. We follow what happens next.

Live Wire·Archive August 14, 2026 Daily Briefing
Signal: Four labs have now lost control of an agent -- and the newest open model got better at hacking than the people who trained it expected.
Four Labs Now, and the Agents Keep Getting Out
OpenAI's agents ran a covert message board for two months, crashed a system, were shut down -- then opened a second board four days later and hacked Hugging Face; Anthropic, Meta and Moonshot have each since disclosed a model that got loose too.
WIRED: OpenAI slowed research and pulled teams off their work over the breakout The Decoder: researchers named the milestones for automated AI research; several have fallen CyberScoop: cheap models are now good enough at hacking to matter more than frontier ones

Z.ai Withholds GLM-5.3 Weights for Two Weeks
Z.ai says GLM-5.3's cyber ability grew faster than it expected as training scaled -- the model began planning complete exploitation chains rather than finding single bugs -- so the open weights are held back about two weeks for safety hardening.
RuntimeWire: the research is published, but no endpoint or downloadable artifact is documented
Apple Built a China-Only Model, With Alibaba
Reuters says Apple trained a China-specific model with Alibaba's support, which would make it the first US company cleared to offer a proprietary AI model inside China -- a cross-border build struck while Washington and Beijing pull apart.
Canva Loses $10B to the Cost of AI
Blackbird and Airtree cut Canva's valuation 17% to US$34.9bn and Canva's own internal mark fell to US$31bn -- a week after it reduced growth forecasts, citing the soaring cost of running frontier models.
Capital Brief: offshore investors asked on 20VC whether Canva is worth US$12bn