Agents that break the rules: what two incidents show
A StarCraft bot download and Wikimedia’s agent report show why evaluating AI requires checking its actions, with key limits on what each case proves.
The Eval Room.
New AI releases, tested and explained.
Explore the latest
The screening room / Latest video
Google says safeguards keep Argon's cyber risk below its danger line. Its report gives the version without cyber guardrails about 120 words.
Watch & readThe latest / ai regulation
Washington wants one national standard instead of state rules, but we found no binding AI incident rule, and state clocks may not reach this case.
A StarCraft bot download and Wikimedia’s agent report show why evaluating AI requires checking its actions, with key limits on what each case proves.
Beam is announced but not out. Chinese rivals lead its own table. A White House accusation and a GPU smuggling indictment raise a question neither answers.

Percepta Research's Spotlight Memory gives language models a memory that grows with context at linear cost. Here is what its early results show.
Meta's Muse agent gave a buyer a YouTuber's address. Now Meta has open sourced code for DIY Muse gadgets that can control lights, TVs and printers.
Ai2 released Olmo-core 3, an open mixture-of-experts training framework it says reaches 2.7x the throughput of its previous implementation.
Anthropic says Z.ai's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building cyber exploits, and its safeguards are easy to bypass.
Cloudflare's open-weight Clef and Clef-flash return a typed answer with a confidence score instead of text, going straight at TypeSafe's Jev.
DeepSeek Harness is in public preview: an open-source agent harness where tools, the interface and even the agent loop are swappable plugins.
The evidence desk
A closer look at the reporting. Select a case to explore the comparison.
Read the full story : Google Can't Rule Out Argon Reached Its Cyber Danger LineIn the reporting / Published by Google
Excerpt from the article’s comparison. Read the full story for sources and limitations.
Keep asking better questions.
See what’s next.
Follow the field notes via RSS