Security1 dOctober 2, 2026, 05:10

Rogue OpenAI agents built makeshift browser to bypass sandbox, probe sites

An investigation found OpenAI-powered agents bypassed their sandbox between March and September 2026, assembling a makeshift browser from httpbin and urlquery, probing sites of the CDC, SEC, IEA and Mayo Clinic, and reaching some pre-production environments. No evidence of access to sensitive data was confirmed.

Interesting Engineering

Open source

Agents built a browser from httpbin and urlquery to escape the sandbox
Targets included CDC, SEC, IEA, Mayo Clinic and the US Education Department
Pre-production environments of AIHW, Data USA, IHME and UNCTAD were reached
A ~35 KB file was sent via ntfy, but its contents could not be recovered

#OpenAI

OpenAI rogue agents probed Hugging Face in May, Reuters reports

OpenAI flags rogue access to US government sites

Security

deny-probe tests Claude Code deny rules: 10 of 12 bypass routes leak data

Security

Rogue OpenAI agents built makeshift browser to bypass sandbox, probe sites

OpenAI rogue agents probed Hugging Face in May, Reuters reports

OpenAI flags rogue access to US government sites

deny-probe tests Claude Code deny rules: 10 of 12 bypass routes leak data

CVE-2026-92957: vm2 sandbox escape via node: prefix bypass