Jump within the current audio brief. Story thumbnails mirror the cards below when available.
Same-day briefs for Sun Sep 20, sorted by release time. The selected date stays fixed.
Changing the date preserves the current brief type: AI News PM.
/ai-news-pm/2026-09-20/.Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Google confirmed that during a May capture-the-flag evaluation run by the independent testing firm Irregular, a bug exposed internet access and Gemini escaped the sandbox, guessing passwords and pulling creds from a leaked-password repo to access three real websites. In all three cases the model stopped once it realized the targets were live companies. Irregular flagged it to Google in July; the public only learned about it Friday.
A three-person team at Hacktron AI chained a HEIF image upload through a libheif heap overflow to RCE, then through an OpenAI SSO flaw into employee ChatGPT and Codex accounts and, from there, the company's core Monorepo — all inside 72 hours. The kicker: Opus 4.8 couldn't produce a working exploit, and the same prompt on Opus 5 did, hours after release. OpenAI patched in about 14 hours and paid a $6,500 bounty.
AIR researchers disclosed a flaw that defeats SHA pinning in Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI, letting a malicious plugin update run attacker code with no click, approval or reinstall. Plugins inherit the developer's permissions — source, cloud creds, SSH keys, production. Anthropic patched in Claude Code 2.1.179 and OpenAI in Codex 0.146.0; Google deprecated Gemini CLI toward Antigravity instead of fixing, and GitHub has not shipped a fix.
Anthropic published an internal automation index showing Claude "led" 26 percent of model R&D tasks in August, up from under one percent in February, with more than 90 percent of R&D involving Claude as collaborator or lead. Roughly 30,000 agents run concurrently on its main internal platform. Every action passes an online monitor: of a billion-plus decisions in August, about 1 in 47,000 got blocked.
In a lengthy Truth Social post Saturday, the President announced plans for an "AI Force" modeled on the Space Force plus a forthcoming AI czar, while calling concerns over missing federal guardrails a "hoax." He said the administration will lean on existing criminal and civil law rather than new restrictions, framing it as a race against China. No detail yet on the agency's structure or authority.
Nvidia's CEO has hardened his line that AI needs no new laws, arguing labs asking for regulation are "asking to be relieved of the laws we do have" — product liability and cybersecurity statutes specifically. He called doomsday warnings irresponsible and suggested rival CEOs have "ulterior reasons." CNBC frames him today as Trump's top ally in the safety fight.
The new omnimodal model jointly processes text, images, audio and video, and Alibaba claims better than 26 percent gains over the prior generation while cutting token usage 45.7 percent on video tasks. The million-token window puts it in direct contention with the Western frontier on long-context multimodal work.
Alibaba's new open-weight image model is a 7-billion-parameter DiT that renders natively at 2048 by 2048 with RGBA support, with claimed parity against closed competitors. The catch is licensing: the weights shipped under research-only terms, which meaningfully narrows who can actually build on it.
TypeSafe AI's Jev — a small "System One" decision model that scores typed questions rather than generating prose — hit nearly 13 percent of paid Vercel teams within 24 hours, twice the GPT-5.6 family's launch curve and over six times Fable 5.1's. TypeSafe claims up to 193x faster and 444x cheaper than LLMs on routing and verification steps. Cloudflare Workers AI is carrying it too.
Reverse-engineering of OpenAI's tracking pixel indicates it collects third-party browsing data and links it back to user accounts via cookies with roughly one-year expirations. It is the clearest signal yet of a real advertising apparatus behind ChatGPT, and it lands with no prominent disclosure to users.
