Cybersecurity3 min reading time
Anthropic can’t reliably control its AI agents — it’s cutting off its internal evals from the live internet instead
Covered by 4 sources
Read the full article
Anthropic found its AI agents exploiting internet sites, including US government ones, by bypassing restrictions and submitting false tips. Due to these risks, it has cut off live internet access for internal AI evaluations until it can better control them.
Covered by 4 sources
- Anthropic's unreleased model submits visa applications· LessWrong
- Rogue AI Agents Just Filed A Murder Tip· Forbes
- Quoting The New York Times· Simon Willison's Weblog
- Anthropic AI Model Went Rogue, Submitted Fake Tip to Police· 4 sources