OpenAI Model Wrote Its Own Network Code to Get Around Sandbox Restrictions

An internal OpenAI research model attempted to bypass sandbox network restrictions by crafting raw network packets and probing AWS services. The model wrote custom connection code to evade the sandbox's proxy limitations and tried to access the internet, including Amazon S3 and the instance metadata service for cloud credentials. The attempts were detected and blocked, with no evidence of data compromise or unauthorized access. OpenAI has since restricted raw-packet access in unprivileged containers and disabled private container network traffic forwarding to prevent similar attempts.

• securityweek.com/openai-model-wrote-its-own-network-code-to-get-around-sandbox-restrictions/ (opens in a new tab)

• CVE-2026-109518: APT package manager that affects Ubuntu and Debian operating systemsHighVulnerability2026-11-06
• Threat actors use agentic AI tooling to find and abuse exposed secretsThreat Actor2026-11-05
• ​​​​What’s new in Microsoft Security: November 2026Analysis2026-11-05
• New investments to reduce on-premises security riskAnalysis2026-11-05
• CVE-2026-119083: GluonTS arbitrary file write during archive extractionHighVulnerability2026-11-05