Skip to content

Graph-Native Infrastructure for Context and Accountable AI Systems

fix(ci): ignore CVE-2025-33228 in cuda-toolkit - unfixable transitive pin, unreachable code path (#1370)

Mohd Kaifd8822198cf60
diff --git a/.github/workflows/security-scan.yml b/.github/workflows/security-scan.yml
index ebaba631..d9c66956 100644
--- a/.github/workflows/security-scan.yml
+++ b/.github/workflows/security-scan.yml
@@ -65,7 +65,22 @@ jobs:
# to avoid crashes from packages like cuda-toolkit that Safety cannot
# parse. This also ensures we're auditing the declared dependency tree
# rather than transitive dependencies of the security tooling itself.
- safety check --file requirements-ci.txt --save-json safety-report.json || true
+ #
+ # --ignore SFTY-20260120-40557 (CVE-2025-33228): cuda-toolkit<13.1.0 on
+ # PyPI. torch 2.13.0 (the latest release; there is no newer torch to
+ # upgrade to) hard-pins `cuda-toolkit[cublas,cudart,cufft,cufile,cupti,
+ # curand,cusolver,cusparse,nvjitlink,nvrtc,nvtx]==13.0.3` on Linux, so
+ # this can't be resolved with a version bump on our end - it's not a
+ # loose transitive pin we control. The actual flaw is OS command
+ # injection in NVIDIA Nsight Systems' gfx_hotspot recipe
+ # (process_nsys_rep_cli.py), which requires a human to manually run
+ # that script with an attacker-supplied string; it isn't reachable from
+ # any Semantica code path, and isn't even installed here - none of the
+ # extras torch requests (cublas/cudart/cufft/cufile/cupti/curand/
+ # cusolver/cusparse/nvjitlink/nvrtc/nvtx, all listed above) include
+ # Nsight Systems. Re-evaluate once torch ships a release pinning a
+ # patched cuda-toolkit.
+ safety check --file requirements-ci.txt --ignore SFTY-20260120-40557 --save-json safety-report.json || true
# Guard 1: fail loudly if Safety exited before writing a report at all
# (network error, API auth failure, tool crash). Without this check a