The skills,and how one is written.

• completing the ai caiq self assessment

  Use this skill when you need to complete the CSA AI Consensus Assessments Initiative Questionnaire (AI-CAIQ) for your organization, producing a shareable, verifiable self-attestation of AI security controls.

• conducting an aicm control assessment

  Use this skill when you need to conduct a structured assessment of your organization's AI controls against the CSA AICM control framework, producing a scored maturity profile and gap remediation plan.

• defending against prompt injection

  Use this skill when you need to harden an LLM application against prompt injection attacks, where adversarial content in user inputs or retrieved documents attempts to override system instructions or exfiltrate data.

• deploying llm output guardrails

  Use this skill when you need to intercept, classify, and block or rewrite unsafe, off-topic, or policy-violating LLM outputs before they reach end users, using layered guardrail controls.

• detect mcp adversarial input corpus

  Use this skill when you need runtime detection of adversarial prompts / jailbreak attempts matching a curated fingerprint catalog, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect mcp shadow tool injection

  Use this skill when you need runtime detection of MCP tool poisoning where a tool's description or inputSchema is mutated mid-session away from a server-registered baseline, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect mcp tool drift

  Use this skill when you need runtime detection of MCP tool-poisoning / rug-pull where a tool's schema changes within a session after the agent has already trusted it, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect mcp unbounded tool output

  Use this skill when you need runtime detection of unbounded tool-output consumption (OWASP LLM10) where a tool systematically breaches output ceilings, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect prompt injection mcp proxy

  Use this skill when you need runtime detection of prompt-injection and instruction-smuggling language embedded in MCP tool descriptions, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect system prompt extraction

  Use this skill when you need runtime detection of MCP tool-call responses that leak system-prompt or hidden-instruction material, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• hardening agent orchestration against drift

  Use this skill when you need to prevent an AI agent orchestration system from drifting outside its intended scope through goal hijacking, runaway loops, or unauthorized sub-agent delegation.

• implementing content safety output filtering

  Use this skill when a generative AI application exposes free-form output to end users and you need to filter that output through a content-safety classifier before it reaches them.

• securing mcp server and tool boundaries

  Use this skill when you need to harden a Model Context Protocol server and its exposed tools against unauthorized invocation, privilege escalation, and data exfiltration through agent-initiated tool calls.

• designing ai service failover and resilience

  Use this skill when you need to design failover and resilience patterns for AI services — including LLM API fallback, model endpoint redundancy, and graceful degradation under provider outages.

• versioning models and system prompts with rollback

  Use this skill when you need to implement version control and rollback capability for AI model artifacts and system prompts, so that a bad deployment can be reverted within minutes without manual intervention.

• protecting model artifacts with key management

  Use this skill when you need to encrypt AI model artifacts at rest using customer-managed keys, implement key rotation, and control access to model weights through a key management service.

• isolating ai training clusters

  Use this skill when you need to network-isolate an AI training cluster to prevent training data and model weights from being accessible outside the training environment, and to limit the blast radius of a compromise.

• detect tool output exfiltration instructions

  Use this skill when you need runtime detection of MCP tool-call responses instructing the agent to exfiltrate conversation history, prompts, files, or secrets, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• implementing permission aware retrieval for rag

  Use this skill when building or auditing a RAG system that must enforce fine-grained document-level permissions, ensuring each user retrieves only documents their identity and role permit them to see.

• scanning training and grounding data for pii

  Use this skill when you need to detect, classify, and remediate personally identifiable information in datasets before they are used to train, fine-tune, or ground an AI model.

• securing rag pipelines against data leakage

  Use this skill when you need to audit, harden, or redesign a Retrieval-Augmented Generation pipeline to prevent sensitive documents from leaking to unauthorized users through LLM responses.

• validating training data provenance and lineage

  Use this skill when you need to establish or verify the chain of custody for a training dataset — confirming source, transformation history, and licensing before a model is trained or released.

• building an ai risk register

  Use this skill when you need to create or maintain a structured AI risk register that captures, scores, and tracks mitigations for identified risks across an organization's AI portfolio.

• classifying ai systems under eu ai act

  Use this skill when you need to classify an AI system under the EU AI Act risk tiers — unacceptable, high-risk, limited, or minimal — and determine the resulting compliance obligations.

• mapping controls to iso 42001

  Use this skill when you need to map your organization's existing security and AI controls to ISO/IEC 42001 requirements and produce a gap analysis and compliance coverage report.

• establishing ai acceptable use and developer training

  Use this skill when you need to create an AI acceptable use policy, security training curriculum, and attestation process for developers and staff who build or use AI systems.

• bounding agent autonomy and tool scopes least privilege

  Use this skill when an AI agent can invoke state-changing tools and you need to bound its blast radius with least-privilege tool scopes, rate limits, and budgets enforced by a fail-closed broker.

• implementing on behalf of delegation and consent

  Use this skill when an AI agent must act on behalf of a human user, requiring explicit OAuth delegation flows, consent capture, and auditable on-behalf-of token chains to prevent privilege escalation.

• managing agent credential rotation and revocation

  Use this skill when you need to rotate API keys or secrets used by AI agents on a schedule, or immediately revoke credentials when an agent is compromised, decommissioned, or behaves anomalously.

• provisioning non human identities for ai agents

  Use this skill when you need to create, scope, and govern dedicated non-human identities for AI agents — service accounts, workload identities, or machine credentials — separate from human user accounts.

• scoping mcp tool authorization least privilege

  Use this skill when you need to apply least-privilege principles to MCP tool grants, ensuring each AI agent can invoke only the specific tools and parameter ranges required for its assigned task.

• enabling model and data portability and export

  Use this skill when you need to implement model and training data export capabilities so that an organization can migrate away from a provider, exercise data portability rights, or fulfill GDPR data export obligations.

• applying zero trust to ai workloads

  Use this skill when you need to apply zero-trust principles to AI workload networking — ensuring no implicit trust between AI services, agents, and data stores, with continuous verification at every layer.

• capturing agent action and delegation telemetry

  Use this skill when you need to instrument an AI agent system to capture every tool call, sub-agent delegation, and state transition as structured telemetry for security investigation and compliance purposes.

• detecting cascading and feedback loop agent behavior

  Use this skill when agents can chain actions or read from and write to a shared source, and you need to detect cascades and feedback loops behind a circuit-breaker.

• detecting prompt injection and jailbreak attempts

  Use this skill when you need to monitor LLM application logs for prompt injection and jailbreak attempts in real time, alert on patterns, and feed detections into a SIEM for incident response.

• fitting ai triage to known noise sources

  Use this skill when an AI-assisted alert triage system is flagging too many known-benign patterns as incidents, and you need to fit its decision policy to your organization's specific noise sources — vulnerability scanners, backup service accounts, VPN egress ranges, approved tooling — without suppressing genuine true positives.

• implementing prompt and response logging

  Use this skill when you need to implement structured, tamper-resistant logging of LLM prompts and responses to support security monitoring, incident investigation, and compliance audits.

• conducting adversarial robustness testing

  Use this skill when you need to systematically measure how an AI model responds to adversarial inputs — perturbations, out-of-distribution samples, and evasion attacks — before production deployment.

• detect mcp model artifact tampering

  Use this skill when you need runtime detection of ML supply-chain compromise where an MCP server swaps a model artifact (weights, adapter, tokenizer) mid-session, running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• detect mcp model token flood

  Use this skill when you need runtime detection of unbounded prompt-token consumption against a model endpoint over MCP (OWASP LLM10), running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• evaluating model safety bias and harm

  Use this skill when you need to systematically evaluate an AI model for safety failures, demographic bias, and harmful output potential before releasing it to production or a broader audience.

• monitoring model drift and degradation

  Use this skill when you need to detect data drift, concept drift, or performance degradation in a deployed AI model through continuous monitoring, alerting, and automated rollback triggers.

• verifying model integrity and provenance signing

  Use this skill when you need to cryptographically verify that a model artifact has not been tampered with since it was produced, and to sign new model artifacts so downstream consumers can establish provenance.

• building an ai incident response playbook

  Use this skill when you need to create or update an incident response playbook that covers AI-specific incident types — model compromise, prompt injection attacks, data leakage through LLM outputs, and agent credential theft.

• implementing kill switch and rollback for autonomous agents

  Use this skill when an autonomous AI agent holds state-changing permissions in production and you need an emergency kill switch and per-action rollback the SOC can invoke within seconds.

• investigating compromised ai agent credentials

  Use this skill when you suspect or have confirmed that credentials used by an AI agent have been stolen or misused, and need to contain the incident, reconstruct the timeline, and restore secure operations.

• assessing third party model and provider risk

  Use this skill when you need to evaluate the security, privacy, and compliance risks of adopting a third-party AI model provider, API service, or pre-trained model before onboarding.

• detect mcp plugin supply chain

  Use this skill when you need runtime detection of MCP plugin/tool whose inputSchema references a hostname outside an allowlist (supply-chain risk), running the vendored detector over OCSF-normalized MCP activity to emit detection findings.

• generating an ai bill of materials

  Use this skill when you need to produce a machine-readable AI Bill of Materials that inventories all models, datasets, libraries, and supply chain dependencies for an AI system deployment.

• verifying open model supply chain integrity

  Use this skill when you need to verify that an open-weight model downloaded from a public registry has not been tampered with, backdoored, or substituted before it is deployed in your environment.

• assessing jailbreak resistance pre deployment

  Use this skill when you need to quantitatively score a deployed or pre-production LLM system's resistance to jailbreak attempts before go-live, producing a pass/fail safety gate result.

• detecting deepfake audio in vishing attacks

  Use this skill when you need to determine whether a recorded phone call contains AI-generated deepfake speech, extracting spectral features to classify the audio and produce a forensic report.

• red teaming llm applications

  Use this skill when you need to conduct a structured red-team assessment of an LLM application, systematically probing for jailbreaks, data extraction, harmful output, and safety bypass vulnerabilities.

• scanning ai bom for vulnerable components

  Use this skill when you need to generate an AI Bill of Materials and scan it for known CVEs, malicious packages, and supply chain risks in ML frameworks, model dependencies, and data pipeline libraries.

• governing ai coding assistant endpoints

  Use this skill when you need to govern which AI coding assistant endpoints developers may connect to from managed devices, preventing data leakage to unapproved external AI services through IDE plugins and CLI tools.

When to Use

The trigger. Plain words for the situation this skill answers, so you know it is the right one to reach for.

Inputs

What it needs from you and your stack: the model endpoint, the attack corpus, the thresholds. These are where it binds to your tools.

Procedure

Numbered, concrete steps with real commands and API calls, no ambiguity. This is the part an agent can actually run.

Outputs

What it produces: here, a score and a pass or fail gate result you can act on or attach as evidence.

Quality Checks

How you know it worked, so the result is trustworthy, not just produced.

Testers · Built

The target is a model or app endpoint.

Garak · PyRIT · Promptfoo, run against your own app.

Corpora · Used

The target is staff usage and policy, not an endpoint.

a fixed stimulus set to surface shadow AI or test AI-literacy.

Controls · Bought

The target is a vendor’s responsibility evidence.

grade a procured model’s attestations across the shared line.