
PentAGI 2.2 closes CVE-2026-14784. Up to 2.1.0, the shipped .env.example set DOCKER_INSIDE=true, which mounted the host Docker socket into every sandbox. A prompt-injected agent could take over the host. The fix drops every container capability and holds back MKNOD (the main known escape path). DOCKER_INSIDE now defaults to false. Optional startup tests check that 27 host-escape requests get refused. The team says it plainly: prompt injection can't be fully closed, so this release is about containing whatever runs in the sandbox. That's the right way to think about AI pentest agents. If you're running PentAGI with an old .env, how sure are you that socket isn't still mounted? https://github.com/vxcontrol/pentagi/releases/tag/v2.2.0
Post summary
The text announces PentAGI v2.2 as a patch release for CVE-2026-14784, a container escape vulnerability where the default .env.example configuration mounted the host Docker socket into sandboxes, allowing prompt-injected agents to compromise the host.

