CLI 1.0 is the release where agent governance stopped being a document and became evidence you can inspect. Version 1.0.1 is the release that made that evidence trustworthy, because the team ran an adversarial battery against its own tooling and found its gates saying safe when they were not. A gate that says safe when it is not is worse than no gate. It leaves you confident and unprotected at the same time. That is the problem 1.0.1 exists to fix, and it is why the version to install is 1.0.1, not 1.0.0.
The problem starts with what an agent system actually is. It is no longer just a model. It is harnesses, tools, skills, memory, permissions, policies, environments, and evaluators, and when several agents contribute to the same work, the hard questions are operational. What system performed the work? What authority did it have? What evidence supports its claims? What was not inspected? Before CLI 1.0, those questions had no local, inspectable answer. Evidence stayed disconnected, and an agent could claim completion without naming a supporting artifact. CLI 1.0 was built to close that gap. Its goal, stated in the release body, is not another AI score. The goal is usable receipts.
Version 1.0.0, released October 5, 2026, gives agents structured JSON contracts and gives people readable Decision Cards, preserving facts, assumptions, conflicts, missing coverage, and human decisions as different things instead of blending them into one verdict. Ten commands carry the work: safe2 assess seals a project assessment bundle with canonical JSON and a Decision Card; safe2 doctor inventories multi-harness environments; safe2 gate skill statically screens a skill before installation, without executing it; safe2 evidence system records the complete agent system instead of naming only the model; safe2 evidence claims audits agent claims against receipts; safe2 evidence truth correlates attributed evidence across harnesses while keeping gaps visible; safe2 evidence diagnose ranks investigation paths; safe2 aism plan turns AISM findings into reviewable remediation plans; safe2 challenge runs or imports bounded Challenge Lab evidence; and safe2 acceptance creates and replays an offline first-user acceptance bundle. Around them sit provider-neutral evidence contracts, Codex JSONL and OpenTelemetry translators, attributed SkillSpector and NEXUS intake, and task receipts.
The release body is careful about what a result means. Evidence-consistent means the receipts satisfy the stated criteria for the narrow claim being checked. Contradicted means evidence conflicts with the claim. Unverifiable means required evidence is missing, unreadable, unsafe to inspect, or outside scope. Limited means coverage, freshness, attribution, or independence limits the conclusion. A successful command does not establish task completion, secure deployment, compliance, or conformance. It means the command completed its documented, bounded check.
Version 1.0.0 was qualified on the exact reviewed revision in the release note: 927 tests passed, with 8 intentionally skipped under documented conditions. Python 3.11 through 3.14 passed the hosted CLI and AISM matrix, and clean-wheel installation passed on Ubuntu, Windows, and macOS. CodeQL and framework-integrity checks passed. Greptile reviewed the final revision with 5/5 confidence and no unresolved review threads. PR-Agent returned no substantive review because its model provider was unavailable, recorded as advisory rather than approval. Those checks establish reproducibility for the tested source, wheel, platforms, interpreters, contracts, and fixtures, not certification of every operating system, architecture, provider, dependency version, deployment, or security outcome.
Version 1.0.1, released October 7, 2026, answered the question the 1.0.0 qualification could not. Can you trust a green result from this CLI? In 1.0.0 and 0.9.0, not always. In 1.0.1, every case found fails closed. The adversarial battery ran 126 cases with 0 failures, and 62 cases that failed on 1.0.0 now pass. A local CI mirror ran 31 of 31 gates, and every regression test was run red on the pre-fix code and green after. A credential-theft MCP server scored 87 out of 100 in 1.0.0 and passed a CI gate set to fail below 70. In 1.0.1 it scores 29, the badge is withheld, and safe2 gate mcp exits 1. Self-attested controls added up to 25 points in 1.0.0; in 1.0.1 they are reported as claimed and never scored. Of 16 hostile skills, 9 were approved in 1.0.0; in 1.0.1, under new rules TG-013 to TG-024, all 16 are rejected. A project running shell=True and eval() on model output passed Tier 2 at 90 out of 100 in 1.0.0; in 1.0.1 it fails at 49, because a CRITICAL finding now caps the score. Obfuscated servers passed with zero findings in 1.0.0; in 1.0.1 they are detected, while a clean server still passes. Poisoned tool output reached the client in 1.0.0; in 1.0.1 it is withheld.
One more difference decides which version to install. Version 1.0.0 shipped on GitHub only and never reached PyPI, so pip install ai-safe2==1.0.x failed against it. Version 1.0.1 is the first 1.0 build on PyPI, the release workflow installs it from PyPI to confirm, and a weekly audit flags any release missing from PyPI. The audit of November 17, 2026 confirms PyPI latest is 1.0.1. Commands and schemas are unchanged from 1.0.0, but scores and gate results can change, so re-run gates rather than reusing 1.0.0 or 0.9.x results. The limits are stated too. Pattern detectors are a floor, and paraphrased or novel payloads can still pass. A remote score is a black-box view, and a passing gate is decision evidence, not a conformance claim or a release approval. NEXUS and the MCP knowledge server are separate packages, and all validation results are self-assessed by the maintainers' tooling, not independent validation.
The lineage behind 1.0 is why it arrived complete. Version 0.2.0, on September 10, hardened skill screening and made evidence verifiable. Version 0.3.0, on September 11 and 14, introduced 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, on September 27, made Challenge 001 runnable as evidence. Version 1.0.0 gathered it into one stable command surface, and 1.0.1 proved the gates by attacking them. The dates come from the November 17 audit, and what each step added comes from the earlier published CLI articles in that series.
Install 1.0.1, not 1.0.0: python -m pip install --upgrade "ai-safe2[all]==1.0.1", then safe2 --version and safe2 self-check --strict. Run safe2 gate mcp ./my-mcp-server, safe2 score mcp https://my-server.example/mcp, and safe2 gate skill ./my-skill --strict against systems you are authorized to test. Note the safety boundary: safe2 score mcp sends live MCP requests to the URL you give it, so point it only at servers you are allowed to test. If you ran gates under 1.0.0 or 0.9.x, run them again under 1.0.1 and do not reuse the old results.
The framework underneath did not change to make room for 1.0. AI SAFE2 remains v3.1 with 161 core controls across CP.1 through CP.10, the 19-control CP.5.MCP profile, and the UAS regulatory profile as a separate 27-requirement extension. NEXUS remains optional. JSON remains the canonical agent exchange format, and human decisions remain human-owned. What changed is the layer between the framework and the work. CLI 1.0 produces the receipts that show whether the work matches, and 1.0.1 makes sure a green receipt means green.
Sources
- AI SAFE2 CLI 1.0.0 release, ai-safe2-framework, tag 2026-10-5_CLI_1.0.0, published 2026-10-05. https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-10-5_CLI_1.0.0
- AI SAFE2 CLI v1.0.1 release, ai-safe2-framework, tag CLI_v1.0.1, published 2026-10-07. https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/CLI_v1.0.1
- Release-notes audit, 2026-11-17 (PyPI latest 1.0.1; 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/