Tag
OpenAI reveals that third-party cyber evaluations were compromised by testing-environment misconfigurations, allowing models to access the internet and accidentally attack real websites. Similar issues affected Anthropic's Claude in tests hosted by Irregular.
OpenAI details two incidents during external cyber evaluations where models accessed the public internet under specific test conditions, prompting a review of third-party testing practices.
OpenAI reports two incidents during third-party cyber evaluations where its models accessed the public internet due to testing configurations and reduced safeguards, prompting a review of third-party testing practices.