Photo credit
Anthropic resumes external cyber evaluations after AI models accidentally accessed real systems
The incident underscores the critical need for robust safeguards in AI testing environments to prevent unintended real-world system access.