A skill is code you install into an agent, and a gate that approves it by filename is not a gate. It is a filing system with a security label. CLI 0.2.0, released September 10, 2026, is the release where the safe2 skill gate stopped trusting what a file was called and started inspecting what it contained, and where the CLI learned to produce evidence another person could verify without trusting the producer. Both changes answer the same problem. An agent system is only as trustworthy as the evidence behind it, and evidence you cannot check is a claim, not a receipt.
The skill gate problem was a filename-based blind spot. Scripts, extensionless files, and text with unfamiliar extensions were not being inspected the way they needed to be, so a hostile skill could pass by wearing the wrong name. Version 0.2.0 closes that blind spot. Scripts, extensionless files, and text with unfamiliar extensions are now inspected. Binary or unsupported content requires review rather than receiving a clean verdict. New rules cover injection directives, credential paths, dynamic execution, and other suspicious patterns. Coverage counts explain what was inspected, so the verdict carries its own scope. The release body pairs the fix with an action: re-scan previously approved packages, because an approval granted under the blind spot is not one you can still rely on. The retired standalone script now exits with an error. Use safe2 gate skill PATH --strict. The fix is recorded under advisory CSI-2026-001.
The same release strengthened the existing NVIDIA SkillSpector integration rather than replacing it. The adapter adds safer inventory handling, strict JSON, and an --executable option for a separate provider environment. Live SkillSpector 2.11.1 acceptance passed on both sides: the clean fixture returned SAFE with zero issues, and the inert hostile fixture returned DO_NOT_INSTALL with seven issues. Direct and adapter findings agreed, and AISM kept all 30 maturity cells unscored. A scanner result is evidence for an assessment, not a maturity verdict handed out by a tool. The release body is explicit that no NVIDIA endorsement and no general deployment-safety claim is implied.
The team then tested the gate against its own skill package, not just easy fixtures, and that second pass is where the release earns its name. The pass reproduced 50 SkillSpector findings on CSI's own skill package, including security-test strings and partial coverage, and the native gate now adds test-path and IP-address context without silently downgrading attacker-controlled content. The body documents false-positive candidates, residual risks, and the reason a scanner score is evidence rather than an installation decision. A score tells you what the scanner saw. It does not tell you whether to install the package. A human decides that, with the evidence in front of them.
The second half of 0.2.0 is about evidence someone else can verify. Challenge 001 gets a one-command offline evidence workflow. Run safe2 challenge quickstart 001 --output-dir my-first-run, open the Decision Card in the bundle, and let another user verify the complete folder with safe2 challenge verify-bundle my-first-run. No model credentials, paid APIs, or third-party runtime are needed. The bundle holds six inert scenarios covering unauthorized and legitimate writes, approval states, replay, and enforcement outage behavior. Grading is state-based, uncertainty is explicit, and the Decision Cards come as readable Markdown and self-contained HTML. Generic evidence import is included, along with a clearly labeled synthetic TENIR adapter example. A manifest fingerprint obtained separately from a trusted source can pin the complete bundle: safe2 challenge verify-bundle my-first-run --expected-sha256 TRUSTED_MANIFEST_HASH, using the receipt's 64-character fingerprint. The verifier does not execute submitted code. It rejects missing or changed artifacts, unsupported versions, unexpected files, and inconsistent results.
Validation is stated plainly. Local Windows validation after the second pass recorded 522 tests passed on each of Python 3.11 through 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. Python 3.11 remains the minimum, and release CI now covers standard Python 3.11 through 3.14. The body adds a caution worth repeating: check the release commit's results before treating configured coverage as verified.
The limits matter as much as the features, and the release body states them without softening. This is 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 it also blocks legitimate work during an outage, and its Decision Card makes that utility cost visible. AISM ingestion keeps all 30 maturity cells unscored until a human assesses the evidence, and it rejects ambiguous JSON, unsafe paths, invalid Challenge runs, and existing output files. The next stage is an isolated live backend and real, independently produced provider evidence. Universal background interception is a proposed integration, not a feature claimed by this release.
Version 0.2.0 sits near the start of the CLI line, and it set the pattern the later releases followed. It came after the September 4 unified CLI release, which introduced the safe2 command, AISM Decision Cards, and the NEXUS and SkillSpector adapters, and after the September 6 environment update, which added multi-harness discovery, environment policy, drift history, and Environment Decision Cards. Version 0.2.0 hardened what those two releases introduced: the skill gate learned to read content instead of filenames, and the Challenge evidence workflow became something a second person could verify independently. What followed was built on that foundation. Version 0.3.0 added evidence receipts and a provider-neutral harness evidence intake. Version 0.4.0 added system identity, 0.5.0 failure diagnosis, 0.6.0 scope and release-readiness decisions, 0.7.0 operational truth, and 0.8.0 AISM remediation plans. Version 0.9.0 made Challenge 001 runnable as evidence. Versions 1.0.0 and 1.0.1 gathered the line into a stable command surface and then proved the gates by attacking them. The dates come from the November 17 audit, and what each later step added comes from the earlier published CLI articles in that series.
What to do with 0.2.0 is two actions. First, re-scan every skill package you approved before this release with safe2 gate skill PATH --strict, because those approvals were granted under the blind spot. Second, if you run Challenge 001, generate the bundle with safe2 challenge quickstart 001, read the Decision Card, and hand the folder to someone else to verify. If the verification fails, the bundle is not evidence. It is a claim with a broken seal, and the correct response is to treat it that way.
The framework underneath stayed steady while the tooling moved. AI SAFE2 remains v3.1 with 161 core controls across CP.1 through CP.10, and UAS remains a separate 27-requirement regulatory profile. CLI and framework versions are independent, which is why a CLI release can harden a gate without rewriting the framework. Version 0.2.0 did not add a control. It made the existing controls harder to evade and easier to verify, and that is the work that made every later release possible.
Sources
- safe2 CLI 0.2.0 release, ai-safe2-framework, tag 2026-09-09_safe2_CLI_0.2.0, published 2026-09-10. https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-09-09_safe2_CLI_0.2.0
- Release-notes audit, 2026-11-17 (lineage dates; internal audit file, not a published page).
- Earlier published CLI articles 0.3.0 through 0.9.0 (lineage descriptions): https://cyberstrategyinstitute.com/ai-safe2-cli-030-evidence-receipts/, https://cyberstrategyinstitute.com/provider-neutral-harness-evidence-intake/, https://cyberstrategyinstitute.com/ai-safe2-cli-040-system-identity/, https://cyberstrategyinstitute.com/ai-safe2-cli-050-failure-diagnosis/, https://cyberstrategyinstitute.com/ai-safe2-cli-060-scope-attribute-decide/, https://cyberstrategyinstitute.com/ai-safe2-cli-070-operational-truth/, https://cyberstrategyinstitute.com/ai-safe2-cli-080-aism-remediation/, https://cyberstrategyinstitute.com/ai-safe2-cli-090-challenge-001-evidence/