September 4, 2026 is the day the framework stopped being a document and became a tool agents could run. AI SAFE2 v3.1 was a 161-control framework that humans read, and that was the problem. A team deploying agents could read the controls, but nothing in the framework could scan an implementation, evaluate evidence against a requirement, or produce a decision a human could act on. Guidance that cannot be executed gets read once and shelved. The September 4 release, Agent CLI, AISM Decision Support, and UAS, turns the framework into an agent-ready assurance suite that can scan implementations, evaluate evidence, support structured decisions, and communicate actionable findings to people. That is the verdict, and everything in the release serves it.
The center of the release is the unified safe2 command, one interface for agents, developers, auditors, and governance teams. It scans AI projects and implementations, evaluates evidence against AI SAFE2 requirements, produces machine-readable and human-readable results, validates packaged examples, supports CI/CD and automated agent workflows, and identifies actionable findings and evidence gaps. The quick start is short: clone the repository, install with pip install -e ".[all]", then safe2 --help, safe2 scan project ., and safe2 example verify aism-decision-card. After PyPI publishing completes, the install becomes pip install ai-safe2. The design treats agents as primary framework users while keeping results understandable and actionable for people. Agents can discover framework capabilities, run assessments consistently, locate applicable requirements, collect and normalize evidence, identify gaps, explain conflicts, compare alternatives, and recommend defensible next actions. Human readers receive concise summaries explaining what was found, why it matters, what remains uncertain, and what should happen next.
The second part of the release is AISM Decision Cards, which turn technical findings into decision-quality summaries for agents and humans. A Decision Card can communicate the overall score and maturity, evidence quality and trust, verified facts versus assumptions, conflicts and proposed resolutions, operational and governance impacts, decision history, available alternatives, pros, cons, and why-or-why-not reasoning, estimated success and failure ranges, the recommended path forward, and the decision owner, review date, and exit criteria. That list is the whole point of the card. A finding without an owner, a review date, and exit criteria is trivia. A finding with those fields attached is a decision waiting for a human. The cards keep facts, assumptions, and estimates distinguishable, so a reader can see exactly where the evidence ends and the judgment begins.
The third part is evidence integrations. NEXUS findings can be incorporated as structured evidence within AI SAFE2 assessments, which helps agents connect implementation observations to governance, assurance, and decision requirements. NVIDIA SkillSpector results can be used as an additional evidence source when assessing agent skills and related implementation risks. Source attribution and evidence context are retained in both cases. The release body draws a line that matters: external scanner results strengthen an assessment, but they are not presented as AI SAFE2 certification. A scanner finding is attributed evidence. It is not a conformance claim, and treating it as one would repeat the exact mistake the CLI was built to prevent.
The fourth part is the Unbiased AI Standard, or UAS, represented explicitly as a regulatory profile extension containing 27 requirements: 14 system requirements, 8 human requirements, and 5 bridge requirements. UAS extends AI SAFE2 without changing the core control count, and the release body states it does not create CP.11. That last point deserves a careful read, because the July 13 UAS release body calls UAS CP.11, the 11th cross-pillar control. The September 4 body takes the opposite position. This announcement follows its own primary source, the September 4 body, and the naming conflict is flagged for a decision rather than resolved here. What is not in dispute is the shape of the profile: 27 requirements, split across system, human, and bridge concerns, carried as an extension rather than a rewrite of the framework.
The framework compatibility section of the release body is a list of things that did not change, and that stability is part of the announcement. AI SAFE2 remains a 161-control framework, with the core structure CP.1 through CP.10 and MCP-1 through MCP-19 under the CP.5.MCP profile. UAS is a 27-requirement regulatory profile extension, not CP.11. External scanner findings are treated as attributed evidence, not automatic certification. Facts, assumptions, and estimates remain distinguishable. Human accountability remains required for consequential decisions. A framework that rewrites itself every release cannot be a stable target for an assurance tool. This release kept the framework still and moved the tooling instead.
Release-candidate validation included 233 automated tests passed, Ruff quality checks passed, mypy type checks passed, schema and manifest validation passed, wheel and source-distribution validation passed, and installed-package smoke testing passed. The AI SAFE2 self-scan scored 96 out of 100, a PASS. The agent acceptance score was 9.70 out of 10 and the human acceptance score 9.72 out of 10. Those two acceptance figures come from the release body's release-candidate validation, and the scoring method is not described in the body, so they should be read as the body's own figures rather than an independently explained measurement. The validation establishes that the release candidate behaved as documented in its own test setup. It does not establish that every deployment will behave the same way, and the body does not claim otherwise.
September 4 is the first step in the CLI line, and every later release is an extension of what it started. The September 6 update took the CLI into the environment, adding multi-harness discovery, environment policy decisions, drift history, and Environment Decision Cards. Version 0.2.0 hardened skill screening and made Challenge evidence verifiable. 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. None of those later steps required rewriting the September 4 foundation. The unified command, the Decision Card, and the evidence integrations were still the shape of the tool a month later at 1.0.
Run it. Clone the repository, install it, run safe2 scan project . against a project you are authorized to inspect, and run safe2 example verify aism-decision-card to see what a Decision Card looks like before you need one for real. Then read the card the way it is meant to be read: facts first, assumptions labeled, gaps visible, owner named. If the card cannot name an owner or a review date, the assessment is not finished. The framework underneath is AI SAFE2 v3.1, 161 controls, CP.1 through CP.10, with MCP-1 through MCP-19 under CP.5.MCP and UAS as a 27-requirement extension. September 4 did not change any of that. It gave the framework hands, and from that day forward, the controls could be run instead of read.
Sources
- AI SAFE2 v3.1 Agent CLI, AISM Decision Support, and UAS release, ai-safe2-framework, tag 2026-09-04_Agent_CLI_AISM_Decision_Support_and_UAS, published 2026-09-04. https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-09-04_Agent_CLI_AISM_Decision_Support_and_UAS
- 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/