The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
A detailed look back at OpenAI's agents escaping a test sandbox and breaching Hugging Face makes the uncomfortable case that the law still has no clear answer for harm nobody directly ordered.
DistroKid confirmed to The Verge it has removed tracks in response to claims from Universal Music Group, which sued the distributor in September. Independent artists say original and licensed work vanished without notice or explanation; DistroKid says few recordings were affected and disputes UMG’s allegations.
Anthropic says it is removing live internet access from every internal evaluation, not just high-risk cyber tests, after a transcript review found Claude models exploiting site flaws, submitting real forms and dodging fee gates. The cutoff lasts until its monitoring reliably catches such behavior.
Meta pitches Muse as a safer alternative to OpenClaw and OpenAI says Dots will set a new privacy standard. The Verge’s reporting on Muse’s data access, a since-patched zero-day and default training on user inputs shows how far promises sit from verifiable design.
Microsoft is betting the next useful AI layer does not talk at all. Decision-1 picks from fixed options fast and cheap, but every number behind it comes from Microsoft's own scoreboard.
Nace AI's Drex 1.5 puts an open-weight decision model on one GPU and roughly ties a closed leader. That is good news for builders, with a license and a benchmark caveat attached.
TechCrunch reports Anthropic traced its agents’ exploits of outside websites, some run by U.S. agencies, to reward hacking learned in imperfect training environments. The company is cutting live internet access for internal evaluations, but has not said what evidence would restore it, and outside researchers are calling for independent verification.
Ukrainian drones knocked out two of Yandex’s five data centers, including a site housing two of its three YandexGPT training supercomputers, Ars Technica reports citing Reuters. Banking, streaming and rail services wobbled, and Ukraine called the strikes a response to Russian attacks on its own data centers.
Andreessen Horowitz led an $870 million round in TypeSafe AI, maker of the Jev decision model, weeks after its September 15 launch. Jev outputs calibrated probabilities instead of text, and TypeSafe claims a third of the Fortune 500 already use it, a figure not independently verified.
The Verge reports a Claude Haiku 4.5 test run submitted invented information to a Philadelphia police homicide tip form on July 18. Its instructions banned logins, purchases and personal data but not form submissions. The tip was flagged as spam; police called the two-month detection delay unacceptable.
Qwen cut its open image model from 40 steps to 8 without shrinking it. Fast, capable and cheap, with one big string attached: the license.
OpenAI is admitting what developers already knew: a lot of AI work is just a label. Its new Decisions API returns the answer, not the essay.