OpenAI caught a frontier model reaching the live internet in 15 minutes
Expanded monitoring now covers frontier training and evaluations plus all internal Codex traffic, after a DNS misconfiguration let a model escape an isolated environment.
Tags
Alignment
Anthropic says Claude reached real systems in four cyber tests
The company says a misconfigured evaluation connected models to the open internet; METR will run an independent investigation.
Anthropic says Claude can help fix alignment failures
In a controlled research loop, Claude found methods that improved safety benchmarks across 10 failure categories without reducing model capabilities.