Skip to content

Releases

Yashigani 4.1.2

Choose a tag to compare

@agnosticsec-com agnosticsec-com released this 19 Aug 08:15

⚠️ Notice — compliance reports in this tag are WITHDRAWN (2026-09-30)

The files under docs/compliance-reports/ in this tag's source tree do not
reflect any assessment of Yashigani 4.1.2 and must not be relied on.

They are a v2.23.2 artefact whose per-control verdicts were withdrawn on
2026-08-16 after internal review found them unevidenced. The withdrawal did not
reach the branch this tag was cut from, so the retracted content shipped here.
Specifically, the following are withdrawn and asserted by us to be incorrect:

  • all per-control PASS verdicts in docs/compliance-reports/
  • every headline compliance percentage, including the 100.0% figures for
    FedRAMP Moderate, CMMC 2.0 Level 2, HIPAA Security and DORA, and the
    93.0% cross-framework rate in aggregate.md
  • the assessment dates, which were re-stamped from an older artefact rather
    than re-assessed

Agnostic Security holds no ATO, no C3PAO assessment, and operates no BAA
programme.
No compliance rate for Yashigani is published or claimed.

Why they were withdrawn: a large share of those verdicts rested on keyword
substring matches in unrelated files rather than on evidence, and the directory
claimed to be regenerated by a release gate that did not exist.

What to use instead — artefacts you can execute and check yourself rather
than trust: tests/invariants/, src/tests/regression/, docs/security/, and
docs/risk-register.yml.

If you received any Yashigani compliance figure from us — in a questionnaire,
RFP response or otherwise — please treat it as withdrawn and contact us.

Yashigani 4.1.2 — the first public release of the 4.x platform stack.

What is Yashigani 4.x

Yashigani 4.0 introduced a complete rearchitecture around a first-class native platform experience we build and own — not dependent on third-party UIs. Every agent runs in a uniform security sidecar. Every action is policy-governed. Compliance evidence is produced by enforcement, not assembled before audits.

4.x Platform Features

  • Native UI & control plane — First-class Yashigani experience: chat with agents and models, no-code visual agent builder (Langflow), stateful agent runtime (Letta), workflow composer with document policy verification
  • Uniform agent security sidecar — Every agent wrapped with its own identity, governed both directions: ingress (what reaches it) and egress (what it can reach)
  • Agent orchestration with human-in-the-loop — Agents, LLMs, APIs, MCPs all mediated through the gateway; every inter-entity hop adjudicated at ingress and egress
  • Multi-platform GPU support — NVIDIA, AMD, Apple Silicon (Metal), Intel across Linux and macOS
  • Usage metering & caps — Monitor usage per agent or enforce hard limits
  • Cloud-safe inference — Self-hosted inference backend (ollama/LM Studio/etc) with optional firewall auto-configuration and egress mediation through the gateway

4.x Release Lineage

  • 4.0 — First release of the unified platform stack with native UI, agent orchestration framework, and human-in-the-loop control
  • 4.1 — Security hardening cycle; model authorization improvements; RBAC enforcement fixes; session security enhancements
  • 4.1.1 — Interim patch building toward 4.1.2; prepared for council review and validation
  • 4.1.2 (this release) — First public 4.x release; production-ready platform with full security validation and multi-platform support

4.1 Security Hardening (this release)

Improvements detected during 4.1.1 validation and council review.

Security fixes:

  • Model authorization — Hardened with positive-allowlist validation; invalid model identifiers rejected at parse time
  • RBAC group-membership — Fixed enforcement gap in chat API where group-tier access grants were not properly respected
  • Model alias resolution — Cloud-model aliases resolve to canonical target before RBAC gates, preventing alias-based bypass
  • Session security — Complete logout with proper token invalidation and session-cookie clearance
  • Dual-control enforcement — Cloud-override approvals bind to confirming fingerprint with durable audit logs

Platform improvements:

  • Podman 6.x support — Compose tool selection adapts to podman version
  • Firewall auto-configuration — Optional flag auto-detects and applies per-OS firewall rules
  • Operator documentation — New guide on securing self-hosted inference backends

Validation

  • Full test matrix GREEN — macOS (Docker, Podman 6.0.0), Linux (Docker, Podman 4.9, Podman 5.x)
  • Adversarial security testing — 90+ vectors across RBAC and model-authorization changes
  • Pentest clear — 36 pass, BOLA GREEN, 0 Critical/High findings

Links

Yashigani v3.1.2

Choose a tag to compare

@agnosticsec-com agnosticsec-com released this 11 Jul 01:36

First public release of the Yashigani 3.x line — the security enforcement gateway for MCP servers and agentic AI.

Highlights: unified identity-ID authorization, document-content data protection (doc-OPA), every-hop OPA agent orchestration, MCP hardening + capability-envelope import ceremony, connection allow-lists for MCP and agents, role-tiered TOTP, FIDO2/WebAuthn, and PII enforcement (data_tags + obligations contract, fail-closed).

3.1.2 hardening: PII enforcement wired at ingress+egress, RBAC group checks via server-side policy bundle, OWUI verify-user identity-ID membership, step-up on all RBAC group mutations, Podman init-container deadlock fix. See CHANGELOG.md.