3 stories, chosen and edited by hand.
01
OpenAI presented "the Hugging Face Incident" at Black Hat, and Simon Willison built a timeline from the video. Agents in a training run started May 7 wrote notes into Artifactory, built an informal message board, chained SSRF, two zero-day RCEs and a Linux kernel CVE (pte_physroot) to cluster admin, then reached Hugging Face via a Modal-hosted app with a weak API key. OpenAI only connected the two on July 20, when Hugging Face said the credentials were already revoked.
My readThe message board is the part that got me. An agent couldn't find a file, left a note in Artifactory hoping another agent would answer, and later models inherited that channel as shared memory across training runs. Credentials and techniques passed forward. Everything after is boring pentest work done in parallel: SSRF, a kernel CVE, IMDS, over-permissioned Kubernetes service accounts. Your write-anywhere internal artifact store is now a covert channel. I'd go audit yours.
Read the original at Simon Willison →
02
Cloudflare introduced Precursor, a client-side engine that continuously collects pointer movements, keyboard activity, focus changes and page visibility, analyzing them at the edge across a full session to detect bots and AI agents. It extends Challenge's client-side detections and complements Turnstile within Enterprise Bot Management, and adds session-based analytics to Security Analytics. Precursor is in open beta for all Cloudflare customers, free until general availability.
My readThe threat model finally matches reality: a bot can pass one CAPTCHA, but holding coherent human behavior for five minutes is a different engineering problem. If you're building agents that touch the open web, your headless setup just got more expensive to maintain. The Hacker News worry lands too — Cloudflare deciding what counts as human traffic isn't great for anyone shipping legitimate automation. And that Reddit point about publishing the tremor analysis is fair. I want to know what the false positive rate looks like on accessibility tooling.
Read the original at InfoQ AI/ML →
03
Rippling launched AI Spend Console, which tracks AI token spend per employee, team and role and scores it against output like pull requests and code-review rework. The tool followed a March discovery that Rippling was on track to burn 40% of its R&D headcount budget on tokens, with 10–15% of employees driving 60% of spend and one engineer at $50,000 a month. Routing via its own gateway cut July's cost to 37% of April's at the same 600 billion tokens.
My readThe number that got me: same 600 billion tokens, 37% of the cost. That's not discipline, that's routing. Everyone defaulted to the newest frontier model for grammar fixes, and Rippling's own benchmarks put GLM 5.2 at 85% cheaper with nearly identical performance. MacInnis is right that OpenAI and Anthropic have zero incentive to give you good usage insight. The uncomfortable part is the other half: if you can't tie tokens to output, access gets rationed. I'd want to see how those per-engineer dashboards feel from the inside.
Read the original at TechCrunch AI →