Skip to content

rhoai-security-reviewer

Consensus-based security review system for RHOAI strategy documents (STRATs). Rather than relying on a single reviewer pass, this plugin uses an orchestrator that extracts a mechanical threat surface inventory from the STRAT, then spawns three independent security reviewers running in isolated forked contexts. Each reviewer checks the STRAT against all 39 catalog risk patterns covering authentication, data protection, cryptographic compliance (FIPS 140-3), network security, supply chain, infrastructure, multi-tenant isolation, agentic AI security, MCP security, and upstream component risk. Reviewers also perform creative exploration for risks not covered by the catalog.

The orchestrator synthesizes reviewer findings using confidence tagging based on cross-reviewer agreement: risks found by all three reviewers are HIGH confidence, by two are MEDIUM, and by one are LOW. Severity is resolved by majority vote. The final output includes a verdict (PASS / CONCERNS / FAIL), a full review with consensus findings, and an actionable security requirements file that is automatically attached to the Jira ticket. The system also cross-references findings against a 47-item NFR checklist derived from 422 prior STRAT reviews, and adds idempotency labels to Jira for pipeline integration.

Plugin Details

Pipeline

rhoai-security-reviewer pipeline

Skills

Skill Description Invocable
/strat-security-review Multi-reviewer consensus orchestrator for security review of STRAT documents. Extracts threat surfaces, spawns three independent security-reviewer instances, synthesizes findings with confidence levels, and produces a final verdict (PASS/CONCERNS/FAIL).
/security-reviewer Individual security reviewer that assesses RHOAI strategy documents against 39 catalog patterns covering authentication, data protection, cryptographic compliance, network security, supply chain, and infrastructure. Uses a two-phase discovery-then-filter approach with severity classification.

Installation

/plugin install rhoai-security-reviewer@opendatahub-skills

Architecture

The plugin follows a two-skill architecture with a clear separation of concerns:

  1. strat-security-review (orchestrator): Manages the full review lifecycle. It fetches the STRAT from Jira via Atlassian MCP, performs mechanical threat surface extraction (endpoints, services, data flows, credentials, CRDs, trust boundaries, RBAC, dependencies, agent/MCP surfaces), determines the review tier (Light / Standard / Deep), and spawns three isolated reviewer instances. After synthesis, it writes two output files (full review + requirements), attaches the requirements file to Jira, and adds verdict labels for downstream pipeline idempotency.

  2. security-reviewer (worker): Runs in context: fork isolation on the claude-opus-4-6 model. Each instance performs a two-phase analysis: Phase A discovers all potential concerns by checking every catalog pattern and performing creative exploration. Phase B applies a relevance gate (requiring specific STRAT citations and checking existing controls) and a severity decision tree. Architecture context from .context/ is consulted for Standard/Deep tiers to avoid flagging already-mitigated controls.

The Light tier includes a short-circuit: if the threat surface inventory is entirely empty, the orchestrator skips reviewer spawning and issues a direct PASS verdict. Intermediate files (threat surface, individual reviewer outputs) are preserved for auditability and calibration.