SV-Gap processes untrusted generated RTL by invoking external EDA tools. Treat candidate files as untrusted input and run evaluations in an isolated container or disposable worker. The v0.1 local runner does not provide a security sandbox.

Model generation and RTL evaluation should be separate trust domains. Generate responses in the environment that holds model credentials, then evaluate the saved responses in a network-disabled disposable worker with no inherited credentials, read-only task inputs, a dedicated writable output volume, and resource limits. The container recipe in docs/evaluate-your-model.md demonstrates this separation.

Do not run contributed functional commands on a workstation containing secrets. Public task packs must not contain credentials, proprietary RTL, or unreviewed shell commands.

Do not publish exploit details for active vulnerabilities. Report them privately to project maintainer Shraddha S at shsridhar@nvidia.com.