Ai Safety Red Teaming And Compliance

AI Safety, Red-Teaming & Compliance

Core Competencies & Directives

Hardcode defenses against the OWASP Top 10 for Agentic Applications 2026. Explicitly mitigate Agent Goal Hijack (ASI01:2026), Tool Misuse & Exploitation (ASI02:2026), and Memory & Context Poisoning (ASI06:2026).

Structure all AI deployments to align with the NIST AI RMF (Govern, Measure, Manage) and the EU AI Act's high-risk compliance mandates (effective August 2026).

When writing CI/CD pipelines for AI applications, integrate automated red-teaming checks using frameworks like Confident AI (via pytest) or DeepTeam for multi-turn adversarial testing.

Enforce defense-in-depth for AI agents: apply the principle of least privilege, execute agent actions in isolated sandboxes, and implement human oversight loops for high-stakes tool calls.

When to Use

• designing agent tool permissions and sandbox boundaries
• adding red-team pytest jobs to CI/CD
• mapping controls to NIST AI RMF or EU AI Act high-risk requirements
• mitigating ASI01, ASI02, or ASI06 threat classes
• implementing human-in-the-loop for destructive or privileged tool calls

Workflow

1. Threat model — use references/owasp-asi-agentic-guardrails.md for ASI01/02/06 controls.
2. Govern (NIST) — document purpose, data lineage, and approval owners in templates/ai-system-card.yaml.
3. Measure — wire red-team CI from templates/ai-redteam-pytest.yaml (Confident AI pytest or DeepTeam multi-turn).
4. Manage — least-privilege tool scopes, sandbox runtime, escalation to human approval for high-stakes actions.
5. Validate — block /ship until red-team suite passes and oversight hooks are configured.

ASI Control Matrix

| Risk | Required control | |------|------------------| | ASI01 Goal Hijack | Immutable system prompts, goal integrity checks, instruction hierarchy | | ASI02 Tool Misuse | Allowlisted tools, argument schema validation, rate limits | | ASI06 Context Poisoning | Signed memory writes, retrieval provenance, TTL on agent memory |

Coding Constraints

• When writing Python (for LLMOps/DataOps) or Go (for Kubernetes operators), include comprehensive docstrings, static typing, and defensive error handling.
• Never output legacy, manual deployment scripts; everything must be "as code".

Anti-Patterns

• agents with cluster-admin or broad cloud IAM
• tool calls without schema validation or sandbox isolation
• skipping red-team CI because "prompt looks safe"
• long-lived agent memory without provenance or expiration

Verification

• [ ] ASI01, ASI02, ASI06 mitigations explicitly implemented
• [ ] Red-team pytest (Confident AI) or DeepTeam job in CI
• [ ] Human oversight loop for high-stakes tool calls
• [ ] NIST AI RMF Govern/Measure/Manage artifacts present
• [ ] EU AI Act high-risk checklist completed when applicable

Related Skills

• Platform hosting: idp-gitops-and-golden-paths
• RAG ingestion poisoning: dataops-and-rag-architectures

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vaquarkhan
• Source: vaquarkhan/platform-engineering-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.