·Archive August 11, 2026 15:07 UTC
Listen to the daily briefingAI-generated audio · covers Aug 10 – Aug 11 · available as a podcast 21:36
Signal: The expensive part of attacking software just got cheap; the labs answer with watermarks, badges and guest lists -- and file for the biggest IPO yet.
Twenty Prompts Cracked Zoom. It Used to Take a Nation-State.
Researchers at A Security used fewer than 20 prompts on public AI models to find a Zoom screen-sharing flaw that silently took over any device on a call -- work they say was nation-state-grade until this year, patched Tuesday.
The Verge: Zoom patched it Tuesday across Windows, macOS, Linux, Android and iOS

NEWAnthropic's $965bn IPO Meets the China Question
Targeting September or early October at a $965bn valuation, Anthropic is fielding investor questions about cheap Chinese models, its standing with the Trump administration and the data-centre backlash -- and answering the first by insisting it only competes at the frontier.
OpenAI Trained a Model to Stop Refusing: 1.5% to 95%
GPT-5.6-Cyber completes 95% of exploit-chain and privilege-escalation requests that GPT-5.6 Sol completes 1.5% of -- released to vetted Daybreak Red partners days after OpenAI paused Astra for being too cyber-capable.
The Decoder: it found two Chrome zero-days, now CVE-2026-15903 OpenAI: the announcement, the two tiers and the 95% number
One EU Law, and Claude Now Marks Every Word on Earth
Every new Claude model will weave an invisible mark into its text worldwide -- Anthropic's answer to one European law, applied to every user on earth, and researchers say it is trivially removable.
Sean Goedecke: text watermarks will always be trivial to remove 9to5Mac: iOS 27 code shows Apple authenticating photos against the camera that shot them
Ghost Note Lab-Grown Brains Are Playing Doom