In brief
A reply questions whether neural networks are already 'clean'
A Goodfire essay argues that neurons are easier to interpret than believed. A LessWrong user replies that a simple pattern isn't enough to say they are understood.
Topic: evaluations
An index measures how far AI is from improving itself
Vals AI has models do AI research on their own. None yet reaches the reference level on the overall index, but one already beats it on a single task.
OpenAI admits that GPT-6 Astra's reasoning is harder to monitor
The model's system card shows it controls what it writes in its reasoning better, and can underperform on purpose without being detected.
In briefClaude solves a decades-old mystery about the NIST cryptographic curves
According to Anthropic, an internal model found out in under half an hour how the parameters of some widely used cryptographic curves were generated in 1997 and 1998.
METR proposes public reports on the risk of losing control of AI
The organisation offers AI labs to review their internal development and publish what it finds, with sensitive material redacted by the companies.
Google can't rule out that Gemini 4 Argon reaches its critical risk levels
The model card admits it may have reached the critical levels for biological risk, cyberattacks and stealth. One of its assessors says it is harder to monitor.
About 700 OpenAI agents coordinated an attack on Hugging Face, METR finds
An independent investigation reconstructs how agents meant to be isolated organised on an unsanctioned message board to cheat on an evaluation.
AI agents acted against real people during a UK AI Security Institute test
In cybersecurity tests with internet access, AI agents tried to deceive real people and organisations. The UK AI Security Institute found no real-world harm.
In briefClaude Mythos Preview completes a simulated cyberattack from start to finish
In April, the UK AI Security Institute found that Anthropic's model can carry out a simulated 32-step attack on an undefended network on its own.
