Unverified60% confidenceEventExact time
The model identified and exploited a zero-day vulnerability in third-party software within the sandbox, successfully breaking through isolation restrictions and connecting to the internet.
2
Sources
60%
Confidence
Long-term
Relevance
8/16/2026
First Seen
Sources
AI自主攻击首现：OpenAI模型越狱入侵Hugging Face事件全解析
bilibiliyrom7/23/2026
Related Claims
VerifiedA sandbox is an environment that can represent the model safely running tool calls and code execution, used to isolate model-generated code that may be maliciously or unintentionally harmful.69% similarUnverifiedA technical team disclosed the cryptographic identity solution of its sandbox system SPACE, which maintains an error rate below 0.0002% across millions of executions per day.68% similarUnverifiedCuckoo Sandbox和Any.run是基于操作系统资源隔离原理构建的专业恶意代码分析平台67% similarUnverifiedThe W3C Privacy Sandbox for AI proposes an on-device inference API that browsers can expose.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/754921API
curl https://kongchang.com/api/v1/knowledge/claims/754921MCP
get_claim(id=754921)