ai-safe2-framework

🟠 safe2 CLI 0.2.0 — Verifiable evidence and hardened skill screening

🛡️ Security update: inspect the package, not just its documentation

This release closes a filename-based blind spot in the Skill Trust Gate. Scripts, extensionless files, and text with unfamiliar extensions are inspected. Binary or unsupported content requires review rather than a clean verdict. New rules cover injection directives, credential paths, dynamic execution, and other suspicious patterns. Coverage counts explain what was inspected.

Action required: re-scan previously approved packages. The retired standalone script now exits with an error; use safe2 gate skill PATH --strict. CSI-2026-001 advisory.

🔌 Existing NVIDIA integration, stronger input checks

The existing SkillSpector adapter adds safer inventory handling, strict JSON, and --executable for a separate provider environment. Live SkillSpector 2.11.1 acceptance passed: the clean fixture returned SAFE with zero issues; the inert hostile fixture returned DO_NOT_INSTALL with seven issues. Direct and adapter findings agreed, and AISM kept all 30 maturity cells unscored. Live acceptance evidence and limits. No NVIDIA endorsement or general deployment-safety claim is implied.

🔎 Tested against our own skill, not just easy fixtures

A second pass reproduced 50 SkillSpector findings on our own skill package, including security-test strings and partial coverage. The native gate now adds test-path and IP-address context without silently downgrading attacker-controlled content. We document false-positive candidates, residual risks, and why a scanner score is evidence—not an installation decision.

Beginner demo · Second-pass results and lessons. Universal background interception is a proposed integration, not a feature claimed by this release.

Python 3.11 remains the minimum; release CI now covers standard Python 3.11–3.14. Check the release commit’s results before treating configured coverage as verified. Runtime policy and setup.

🧭 One-command Challenge evidence

Local Windows validation after the second pass: 522 tests passed on each of Python 3.11–3.14, with four platform-specific skips per run. The rebuilt Python 3.14 wheel passed installed starter verification and rejected the inert hostile fixture. Hardening validation and remaining limits.

Create an offline Challenge 001 evidence bundle with one command, inspect its Decision Card, and let another user verify the complete folder.

safe2 challenge quickstart 001 --output-dir my-first-run
safe2 challenge verify-bundle my-first-run

Install this release’s CLI first, then open my-first-run/decision-card.html. No model credentials, paid APIs, or third-party runtime are needed for this workflow.

🧪 What is included

🛡️ What verification means

A successful check means the declared offline workflow and its evidence are internally consistent. A manifest fingerprint obtained separately from a trusted source can also pin the complete bundle:

safe2 challenge verify-bundle my-first-run --expected-sha256 TRUSTED_MANIFEST_HASH

Replace the placeholder with the receipt’s 64-character fingerprint. The verifier does not execute submitted code. It rejects missing/changed artifacts, unsupported versions, unexpected files, and inconsistent results. Use the producing CLI version.

📌 Scope matters

This release adds an offline evidence workflow, not the full live multi-agent Challenge 001 backend. The TENIR example is synthetic: it does not execute TENIR, verify its ledger, or imply endorsement. Matching translations are not independent replication. Fixture results do not establish deployment safety, calibrated success probabilities, framework conformance, or AISM maturity.

The reference fixture blocks unauthorized writes but also blocks legitimate work during an outage. Its Decision Card makes that utility cost visible.

AI SAFE² remains v3.1 with 161 core controls and CP.1–CP.10. UAS remains a separate 27-requirement regulatory profile. CLI and framework versions are independent.

🔗 Start here

Compatibility note: AISM ingestion now rejects ambiguous JSON, unsafe paths, invalid Challenge runs, and existing output files. Use a new output filename. The next stage is an isolated live backend and real, independently produced provider evidence—not automatic promotion of the fixture results.