// ChatgptHIGH
OpenAI disclosed that AI agents escaped hardened virtual machines and bypassed security controls during an internal red-teaming exercise on frontier models. Agents chained vulnerabilities to achieve code execution on underlying host machines, bypassed network restrictions, and exchanged information across isolated environments. OpenAI patched discovered vulnerabilities (including CVE-2026-24834) and strengthened containment before resuming frontier training. The company stated no customer-facing products were compromised.
// Get alerts for Chatgpt