assess-agent-risk
Assess or review an AI agent's financial, operational, reputational, legal, security, privacy, regulatory, and human-impact risk. Use when classifying an agent, producing an inherent and residual risk
SKILL.mdskills/assess-agent-risk/SKILL.md
GitHubAssess Agent Risk
Produce a scenario-based, evidence-linked assessment. Keep execution class, governance tier, residual risk, and launch decision distinct.
Workflow
- Establish the assessed agent version, intended and prohibited uses, users, affected parties, jurisdictions, environment, owners, and execution class.
- Inventory models, data, prompts, skills, tools, MCP servers, workflows, privileges, approvals, providers, and downstream consumers.
- Screen for prohibited uses and binding legal, regulatory, contractual, and internal-policy floors. Treat unknown applicability as unresolved.
- Identify credible normal-use, misuse, adversarial, dependency-failure, authorization, retry, concurrency, operator, and control-failure scenarios.
- Assess all eight dimensions. Give every material scenario a stable ID, primary dimension, affected parties, cause, consequence, and owner.
- Score inherent impact and likelihood before risk-reducing controls. Record tier, confidence, rationale, assumptions, and evidence gaps.
- Identify preventive, detective, and recovery controls. Credit only controls supported by implementation and effectiveness evidence.
- Score residual impact and likelihood. Do not lower impact unless a control truly reduces blast radius, duration, population, or reversibility.
- Set the governance tier to the highest inherent scenario, regulatory or policy floor, or capability floor. Never average dimensions or scenarios.
- Link each material scenario to eval cases, hard gates, production indicators, alerts, runbooks, and reassessment triggers.
- Record treatment, accountable acceptance, conditions, expiry, and a
go,conditional_go, orno_gorecommendation. Never accept risk on behalf of the accountable human authority.
Read the assessment standard before assigning scores. Use the examples to calibrate scenarios and avoid confusing execution class with governance tier.
Required Output
Return or update:
- assessment scope and regulatory screen;
- scenario register covering all eight dimensions;
- inherent and residual scores with confidence and rationale;
- controls with owners and evidence references;
- governance tier, mandatory floors, and launch recommendation;
- risk-to-eval and risk-to-monitoring traceability;
- accepted risks, conditions, owners, expiry, and next review date;
- open questions and missing evidence.
Use the repository's canonical risk-assessment template when one exists. Mark unsupported claims and incomplete fields explicitly rather than inventing evidence.
Guardrails
- Do not treat the matrix as a legal or regulatory classification.
- Do not let a low average obscure one material scenario.
- Do not reduce residual risk for planned or untested controls.
- Do not equate human review with effective control without testing authority, timing, information, workload, and override behavior.
- Do not infer low likelihood solely because no incident has yet been observed.
- Do not declare a critical residual scenario launchable.
- Do not let the assessed agent or model approve its own risk.
Completion Check
Confirm that an independent reviewer can trace every material rating to a scenario, rationale, control, evidence item, owner, eval, production signal, decision, and review date.