AI SAFE2 v3.1 safe2 CLI Update: Assess the Environment the Agent Runs In

The September 6 CLI update is the release where safe2 stopped assessing the project and started assessing the room the project runs in. An agent does not operate in a vacuum. It runs on a workstation with several harnesses installed, inside WSL distributions, against SSH-accessible hosts and cloud VMs, under configuration files that change without announcement. An assessment that looks only at the project misses the second harness, the drifted instruction file, and the configuration change that quietly widened what the agent may do. That is the problem this update was built to solve. It builds on the September 4 Agent CLI release, which introduced the unified safe2 command, AISM Decision Cards, NEXUS and SkillSpector evidence adapters, and the UAS profile.

The headline capability is multi-harness discovery through safe2 doctor, which identifies command and configuration indicators for Codex, Claude Code, Antigravity, Hermes, OpenClaw, and Grok, alongside host, shell, CI, and WSL metadata. Discovery matters because an organization cannot govern harnesses it has not inventoried. The update also adds explicit WSL and SSH targets, so teams can assess metadata from selected WSL distributions and SSH-accessible Linux hosts or cloud VMs. Incomplete or unreachable targets appear as coverage gaps rather than being silently skipped. A gap you can see is a decision. A gap you cannot see is an assumption.

Around discovery sits a project asset inventory that finds agent instructions, skills, MCP configuration candidates, persistent state, workflows, container definitions, and infrastructure files, with optional hashes to support content-change detection. Opt-in configuration inspection collects an allowlisted structural summary of supported local JSON and TOML configurations, including selected approval and sandbox settings, without emitting raw configuration values. The CLI reports the shape of the configuration, not its secrets. Environment posture and policy turn those findings into decisions: review findings and candidate control mappings, then evaluate a named policy, and with enforcement enabled, ALLOW, DENY, and HOLD produce exit codes 0, 1, and 2 for agents and CI.

The update also adds reviewed baselines and drift, comparing assets, harness indicators, configuration hashes, target coverage, and assessment scope against an earlier inventory. Drift is the quiet failure mode of agent governance. An instruction file changes, a configuration hash moves, a harness appears, and the system approved last month is no longer the system running today. Baselines make that visible. Environment Decision Cards then turn the whole picture into a human decision document, in Markdown and HTML, with facts, assumptions, conflicts, history, impacts, alternatives, recommendations, and exit criteria. Operational feedback evidence records sanitized descriptions of false completion, missing evidence, silent failures, context loss, stuck loops, permission friction, and integration failures. Versioned schemas and run manifests assemble artifacts with digests, source contracts, integrity status, and coverage in one run record. An executable environment example reproduces a baseline-to-drift-to-policy workflow, and a successful verification observes policy DENY, one drift change, and zero invalid manifest artifacts.

The update hardens the tooling at the same time. It hardens bearer-token transport and redirect handling in the remote MCP assessor, bounds file reads, scan traversal, provider output, and subprocess execution, improves malformed-evidence handling, and fixes Windows console encoding failures. Relevant CI and publishing actions now use pinned commit SHAs. The NEXUS and SkillSpector adapters emit dedicated contracts, safe2.nexus-evidence.v1 and safe2.skillspector-evidence.v1, and older provider bundles must be regenerated before using the new manifest validation. SkillSpector remains optional and installed separately, with upstream attribution retained.

The boundaries are stated in the release body. Discovery is metadata-only by default. Asset hashing and local configuration inspection are separate opt-ins. WSL and SSH probes inspect explicit targets only. Cloud control-plane inventory and automatic network discovery are outside this collector's scope. SHA-256 seals support integrity comparison, but they are unsigned and do not authenticate the evidence author. AISM scores and environment posture answer different questions, and inventory metadata does not establish implementation maturity. Cards identify unsupported probability estimates as NOT ESTIMABLE, and evidence sources, assumptions, and human ownership remain visible.

The workflow shares the existing safe2 command surface. Project, skill, and MCP assessment run through safe2 scan, safe2 score, safe2 gate, and safe2 report. Environment discovery and decisions run through safe2 doctor. AISM assessment runs through safe2 aism init, ingest, score, and compare. NEXUS and optional SkillSpector evidence run through safe2 evidence nexus and safe2 evidence skillspector. Evidence validation and run records run through safe2 schema and safe2 evidence manifest. Feedback and examples run through safe2 feedback and safe2 example verify. Agents receive structured JSON and predictable policy outcomes. Humans receive a Decision Card explaining what was found, why it matters, what remains uncertain, and what action to take.

Installation needs Python 3.11 or newer, from a fresh repository clone: git clone https://github.com/CyberStrategyInstitute/ai-safe2-framework.git, then python -m pip install -e ".[all]". The core commands are safe2 doctor . --assess --output environment-inventory.json --card-format markdown --card-output environment-card.md, followed by safe2 schema validate discovery-v1 environment-inventory.json and safe2 evidence manifest environment-inventory.json --subject-id local-workstation --output run-manifest.json --strict, then safe2 example verify environment-decision-card. To include a structural configuration review, add --inspect-config to the doctor command.

Validation recorded for the implementation in PR #316 includes 319 local automated tests passed and 20 GitHub checks passed, with CLI and AISM checks on Ubuntu with Python 3.11 and 3.12, and separate NEXUS checks with Python 3.10, 3.11, and 3.12. Ruff, mypy, repository UX, Markdown, schemas, manifests, examples, and framework-invariant checks passed, along with wheel and source builds and installed-package example verification. The strict skill trust gate returned APPROVE. Those are the results for this update. The September 4 release retains its own validation record, and framework conformity and deployment decisions depend on scope and evidence.

The September 6 update is the second step in the CLI line, and the steps that followed kept the same direction: outward from the project into everything around it, then back into evidence a human can verify. 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.

Run this update against the machines where your agents actually work, not against a clean test project. Run safe2 doctor with --assess, review the Decision Card, and treat every coverage gap it names as a question to answer, not a line to ignore. Establish a baseline, then run it again after the next configuration change and read the drift. The framework underneath did not move: AI SAFE2 remains a 161-control framework across five operational pillars plus the Cross-Pillar Governance Layer, with CP.1 through CP.10, the 27-requirement UAS profile as a separate extension, and MCP-1 through MCP-19 inside CP.5.MCP. What moved is the field of view. The CLI now sees the environment, and governance that cannot see the environment cannot govern the agent.

Sources

  • AI SAFE2 v3.1 safe2 CLI update release, ai-safe2-framework, tag 2026-05-09_safe2_CLI_Update, published 2026-09-06. https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-05-09_safe2_CLI_Update
  • 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/

Recent Posts

Tag Cloud

Stop Threats Before They Execute

Your free Kernel-Level Defense Buyer’s Guide is ready to download.

By providing my email address, I consent to receive emails and text messages—including newsletters and marketing communications—from creators of Warden Secure, Cyber Strategy Institute, our flagship zero-trust platform for ransomware prevention, and agree to the Terms and Privacy Policy. You may unsubscribe at any time.