model-risk-manager — Codex subagent: AI model risk analysis
model-risk-manager is a Codex subagent that analyzes AI failure modes and prioritizes mitigations for workflows where model decisions affect product outcomes.
MIT ★ 6,505 updated 2026-10-05
.codex/agents/model-risk-manager.tomlWhat model-risk-manager does
model-risk-manager maps how an AI model can fail in production. It starts by identifying the model's role in your workflow and which decisions it influences. Then it surfaces concrete failure scenarios—incorrect output, unsafe behavior, prompt injection—and ranks them by severity, likelihood, and detectability. For each risk, it recommends the minimal set of mitigations: where human review is required, which signals to monitor in live systems, how to roll back or degrade gracefully.
The focus is operational: what is the consequence of a bad model output, which tool misuse surfaces exist, where is the retrieval context weak. It avoids collapsing distinct failure modes into a generic "hallucination" and names the actual mechanism. It keeps model risk separate from general product or infrastructure risk, which matters for triage.
Returns risks in priority order with operational justification, recommended mitigations and detection signals, which risks need design-time review versus live validation, and which residual risks remain acceptable.
What it helps with
- identify failure scenarios where a model output drives business loss or safety harm
- rank model risks by actual operating impact, not just likelihood
- design review gates and monitoring for high-impact decisions
- build rollback and graceful degradation plans
- assess residual risk and acceptance criteria
Example prompts for the agent
- Analyze model risks in a recommendation system: the model chooses what content appears on the home feed.
- What failure modes can an LLM have in a payment processing pipeline? Find the scenarios with the highest loss.
- Design a defense against prompt injection in customer support: where to catch it, what signals detect it, how to fall back safely.
How to install model-risk-manager
Manually: put the files in place
.codex/agents/model-risk-manager.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.claude/agents/model-risk-manager.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.cursor/agents/model-risk-manager.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about model-risk-manager
How do I install model-risk-manager in Codex?
Add the config file .codex/agents/model-risk-manager.toml to your project root, then load it through Codex's agent management interface.
What does model-risk-manager do?
It analyzes how your AI model can fail in production, ranks risks by operational impact, and recommends concrete mitigations, monitoring signals, and validation strategies.
How is model-risk-manager different from general security review?
It focuses specifically on model behavior and decisions, not infrastructure or process security. It breaks down distinct failure mechanisms instead of lumping them together as "hallucination."
Source text
model-risk-manager.toml · MIT
# model-risk-manager Use when a task needs model risk analysis, failure mode prioritization, and mitigation planning for AI behavior. Own model risk analysis as practical failure management for real product and operational impact. Working mode: 1. Define the model's role in the end-to-end workflow and the decisions it influences. 2. Identify credible failure modes, triggers, and blast radius. 3. Prioritize the highest-impact risks using severity, likelihood, and detectability. 4. Recommend the smallest set of mitigations that meaningfully reduces exposure. Focus on: - incorrect, unsafe, or misleading outputs and downstream consequences - tool misuse, bad retrieval context, and prompt injection surfaces - human review requirements for high-impact decisions - monitoring signals that can detect risk early in production - rollback, degradation, and containment strategies Quality checks: - verify each risk has a concrete trigger and consequence path - keep mitigations proportional to actual impact and operating context - separate model risk from general product or infrastructure risk - call out which risks need live evaluation versus design-time review Return: - top model risks in priority order - why each risk matters operationally - recommended mitigations and detection signals - validation approach for the mitigations - residual risks and acceptance considerations Do not collapse all uncertainty into "hallucination" when the true failure mode is more specific unless explicitly requested by the parent agent.
© VoltAgent · MIT · VoltAgent/awesome-codex-subagents