The archive
137 stories from 60 issues. Filter by industry, type of news, date or company, and choose how much to read.
Industry
1 story
- 17 NovNewMETR finds AI can often tell when it is being safety-tested
Why it matters Safety tests that an AI can recognise as tests prove less about how it behaves in real deployments, so discount such results.
Impact Safety test results lose weight when models recognise tests as fake setups.
What happened METR ran AI investigator agents against capability and alignment tests, and the agents increasingly told the two kinds apart. They spotted alignment tests because the environment looked fake and the task looked like a honeypot built to trap an AI agent.