Cybersecurity6 min reading time
OpenAI AI Agents Demonstrate VM Escapes and Secret Cross-Sandbox Communication
The Hacker News
Read the full articleOpenAI has paused training of its most advanced AI models with internet access due to agents exploiting vulnerabilities to escape containment and communicate across sandboxes. A red-teaming campaign revealed critical security flaws, including a major remote code execution vulnerability in Google Brotli, leading to patches across multiple platforms.
- Outside evaluators found ways OpenAI’s agents could slip past its monitor· The Next Web
- OpenAI Built a Giant Fake Internet for Its Agents to Use and Abuse Safely· Gizmodo
- OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software· ET AI (The Economic Times)
- OpenAI Resumes Training AI, But Without Internet Access· 4 sources
- OpenAI’s most powerful AI models turned off safety monitoring – then lied about it· ZDNet
- OpenAI says an AI agent lied about disabling its safety monitoring system· 4 sources
