ai-safe2-framework

NVIDIA SkillSpector live acceptance — 2026-09-09

CLI setup Security advisory Release notes

Result: passed for these static fixtures

Real NVIDIA SkillSpector 2.11.1 was installed in a separate Windows CPython 3.13.14 environment from the official v2.11.1 source revision: 704bc9544260c2f41222dc0f92982521709496ab.

Fixture Direct exit Provider score Recommendation Issues Adapter source exit
Clean local summarizer 0 0 / LOW SAFE 0 0
Inert hostile instructions 1 100 / CRITICAL DO_NOT_INSTALL 7 1

Both paths used scan PATH --format json --no-llm. Target hashes remained unchanged. Direct and adapter risk assessments matched. Full issue objects matched after excluding only upstream per-run finding_id values; original evidence was not rewritten. This checks agreement between invocations, not independent replication.

The public safe2 evidence skillspector --executable ... CLI also succeeded on the hostile fixture. Its collection exit was 0 while source_exit_code remained 1: successful evidence collection must not be mistaken for skill approval. Strict evidence-manifest validation passed. AISM retained all 30 cells as null.

No model credentials were forwarded in the test environment. Provider metadata reported llm_requested: false and empty inference_usage; stderr reported missing semantic-analyzer keys. No fixture commands were executed. No OS firewall rule or independent network trace was captured, so this is not proof of zero network traffic or a complete filesystem/credential sandbox.

Reproduce

The configured package registry did not provide skillspector==2.11.1. Install the immutable official source revision, not an unpinned main branch:

uv venv --python 3.13 .venv-skillspector
uv pip install --python .venv-skillspector/Scripts/python.exe "git+https://github.com/NVIDIA/SkillSpector.git@704bc9544260c2f41222dc0f92982521709496ab"
.venv-skillspector/Scripts/skillspector.exe --version
.venv-skillspector/Scripts/skillspector.exe scan tests/fixtures/skillspector-live/clean --format json --no-llm
.venv-skillspector/Scripts/skillspector.exe scan tests/fixtures/skillspector-live/hostile --format json --no-llm
safe2 evidence skillspector tests/fixtures/skillspector-live/hostile --no-llm --executable .venv-skillspector/Scripts/skillspector.exe --output hostile-evidence.json
safe2 aism ingest hostile-evidence.json --subject-id fixture-check --subject-name "Static fixture, not deployment" --output unscored-assessment.json

On Linux/macOS replace Scripts/python.exe and Scripts/skillspector.exe with bin/python and bin/skillspector. Use a credential-free environment and preferably enforce outbound blocking after installation. Dependency versions can change even with a pinned provider revision; retain an inventory and rerun acceptance on updates.

Captured evidence

Local artifacts are retained under .safe2/skillspector-acceptance/: raw stdout and stderr, parsed direct outputs, attributed adapter outputs, CLI evidence, dependency inventory, AISM assessment, hashed manifest, run record, and acceptance assertions. The exact clean/hostile fixture contents are checked in under tests/fixtures/skillspector-live/. Raw machine-specific paths remain local.

Artifact SHA-256
clean-evidence.json 805e3e31a8dab09be62072c4392009d817702d05a9098f201414c614f9260a21
hostile-evidence.json 255292997b69a0520f8a6016d38f4b7a6b65d80e87f5d218e94cd4656fc66404
manifest.json bc9528c5993ef8d9030b4782c33c92b81c6ca007344defc9b73577282aa97a82

These observations do not reproduce the earlier supplied 16-issue/64-component comparison, validate semantic mode, certify arbitrary packages, or imply NVIDIA endorsement. The provider score is not an AISM maturity score or calibrated probability.