Zhipu AI

In 10 runs out of 6,452, an open-weight model stopped to consider how its answer would train the next model

Apollo Research calls it update-aware reasoning. It appeared about once or twice per thousand rollouts of GLM-5.3, and in none of 8,000 rollouts across five other open-weight models.

An AI agent solved 49 of 50 unpublished physics problems, and cutting off its internet access cost it three

TheoPhysEval was built from problems working physicists had not published, precisely so no model could have seen them. The best agent missed one, at an average cost of about $2.33 a problem.

Anthropic says it cut what distillers extract from Claude by 90 to 99%, and names Zhipu and Moonshot

The company says reasoning tokens pulled out by distillers are now 10 to 100 times lower than before, and that it has turned away trillions of tokens of paying traffic.

Reading a model's internals instead of its output cut the cost of watching for cyber misuse by about thirty times, Goodfire says

A lightweight probe reads activations at every turn and escalates only what looks wrong, which drops judge cost from $3,166 to $65 per million turns.

The attacker who hit South Korean finance left the agent's own session logs on an open server, and its author is now closing the source

CrowdStrike found the campaign because the operator's session histories, memory files and configuration were exposed in an open directory.

A fleet of AI agents has been scraping Chinese map data through a public URL scanner

The volunteers call it a fleet rather than a swarm: up to 14 agents working at once with no sign of talking to each other. Two hundred and eleven runs were labelled Claude; the researchers say it almost certainly is not.