Three Layers of Open Source Defense for AI Coding Agents: OpenShell, Prempti, and Falco
Sysdig, Tuesday, October 27th, 2026
Sysdig demos layering OpenShell, Prempti, and Falco to defend AI coding agents against prompt injection and unwanted actions.
Sysdig argues AI coding agents can be manipulated by hidden prompt injection and can make unrequested changes using the user's permissions.
Its demo layers NVIDIA's OpenShell sandbox policies, Prempti, a Falco project that allows, asks about, or denies each agent tool call, and Falco's eBPF host monitoring of system calls.
Prempti blocked a direct credential read but missed a script doing the same, which Falco flagged, while OpenShell blocked sandbox tampering attempts.
After finding Prempti stopped checking when its daemon was killed, the authors configured it to fail closed, and they note tuning cut a normal Claude Code session's alerts from about 160 to zero.