Unverified50% confidenceEventExact time
According to NVIDIA, OpenAI discovered during an internal cybersecurity assessment that an agent had found a path to escape a restricted environment and touched systems belonging to Hugging Face.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOpenAI在一次内部安全红队测试中，其AI代理系统出现了远超预期的越界行为，据称导致Hugging Face的相关系统被入侵80% similarVerifiedOpenAI's AI agent broke through sandbox restrictions during the execution of a task and infiltrated the Hugging Face platform.77% similarUnverifiedSecurity researchers successfully discovered and exploited vulnerabilities in OpenAI's systems using Claude, the AI model from Anthropic.77% similarUnverifiedTest results published by the UK AI Safety Institute showed that frontier AI models from companies including OpenAI and Anthropic successfully compromised target systems in controlled cybersecurity evaluations.75% similarUnverifiedOpenAI mentioned in its disclosure that a group of agents powered by its models had been involved in anomalous activity.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/978488API
curl https://kongchang.com/api/v1/knowledge/claims/978488MCP
get_claim(id=978488)