Best AI Agent Skills for Compliance and Audits (2026)

No agent skill will get you through a SOC 2 Type 2 audit. It's worth saying that before anything else, because the marketing around compliance skills implies otherwise, and the gap between "drafted a policy" and "produced twelve months of evidence a CPA firm will accept" is where compliance programs live.

What these skills do is narrower and still worth having. They give an agent the control catalogs, article numbers, and clause references it otherwise invents. They turn a gap assessment from a two-week reading exercise into an afternoon. And a couple of them address a problem most compliance teams haven't priced in yet: the agent skills your engineers installed last month are now inside your audit scope.

Star counts and update dates below come from our skill database as of August 1, 2026.

What's available

All free, all open source.

Framework knowledge: the GRC skill pack

Claude-Skills-Governance-Risk-and-Compliance is 30 separate skills, one per framework, under MIT.

claude skill add Sushegaad/Claude-Skills-Governance-Risk-and-Compliance

The coverage is unusually wide for a community project: ISO 27001, SOC 2, FedRAMP, NIST CSF and 800-53, CMMC 2.0, CIS Controls v8, NIS2, GDPR, HIPAA, CCPA/CPRA, LGPD, India's DPDPA, ISO 27701, PCI DSS, DORA, SWIFT CSP, ITAR and EAR, plus the newer AI standards: ISO 42001, NIST AI RMF, and the EU AI Act. Sustainability and accessibility get a look too, with CSRD, WCAG, and Section 508.

The repo reports its skills scoring 94% against an 81% baseline across 150 test cases and 752 assertions. Read that as the author's own benchmark, evaluated by other agents rather than by auditors, because that's what it is. I checked the repository directly on August 1, 2026 and those are the figures published there. Directory listings elsewhere quote a 72% baseline; that number is stale.

What it changes in practice: ask an agent without this installed for the SOC 2 criteria covering change management and you'll get a confident paraphrase with no criterion ID. Ask with it installed and you get CC8.1 by name. For anyone drafting a policy set or running a gap assessment, that difference is most of the work.

The tier matters here. It's community, unreviewed by us, from a single maintainer. That's fine for a knowledge pack that reads files and writes markdown. It would not be fine for something with network access and credentials.

Control testing: Prowler

Knowledge is the easy half. Evidence that a control is operating is the hard half, and it's the half auditors care about.

Prowler tests live cloud configuration against published benchmarks (CIS, plus the compliance frameworks layered on top of it) across AWS, Azure, GCP, and Kubernetes.

claude skill add prowler-cloud/prowler

At 14,508 stars with commits as of July 31, 2026, it's the most active project on this page. Running it through an agent is a real workflow: scan, get a few hundred findings, and have the agent group them by which framework control they map to and which ones are false positives for your architecture. That triage is tedious for a person and well suited to a model.

Where it stops: Prowler tells you the state of your cloud right now. SOC 2 Type 2 asks whether the control operated throughout the observation window. Point-in-time output is one input to that argument, not the argument.

The part most teams miss: your skills are in scope

Every SKILL.md your team installs is third-party code executing with your engineer's credentials on a machine that touches your source. That's a vendor-risk question and a change-management question at the same time, and most control frameworks already have language covering it even though nobody wrote the language with agent skills in mind.

Two tools address this directly.

SkillSpector scans agent skills for malicious patterns before they run. It comes from NVIDIA, sits at 14,001 stars, and is Apache-2.0 with commits as of July 31, 2026.

uv tool install git+https://github.com/NVIDIA/skillspector.git, then skillspector scan ~/.claude/skills

It checks 68 vulnerability patterns across 17 categories, including prompt injection, data exfiltration, privilege escalation, and supply-chain risks, using static pattern matching plus optional semantic evaluation by a model. NVIDIA's own figures for why this exists: 26.1% of skills they examined contained vulnerabilities and 5.2% showed likely malicious intent. It also runs in Docker, works as an MCP server, and exports SARIF, so the results drop straight into whatever code-scanning pipeline your engineering team already reports from.

That last detail is the compliance-relevant one. Running it across ~/.claude/skills gives you a documented screening step applied to third-party components, in a format your existing evidence process already handles. That's a control, and right now almost nobody has one.

SkillHub goes further and gives you the registry itself: publish and version skill packages internally, gate access with RBAC, keep audit logs, deploy on-premise via Docker or Kubernetes.

claude skill add iflytek/skillhub

If your auditor asks how you control what code runs inside your development environment, "engineers install from GitHub at will" is a finding. A private registry with approval and logging is an answer. For a regulated organization, SkillHub is the highest-value item on this page even though it does no compliance work of its own.

AI-specific risk: SAIL

ISO 42001, the NIST AI RMF, and the EU AI Act all now ask organizations to enumerate AI-specific risks, and most existing risk registers have nothing useful to say about prompt injection or agent tool misuse.

SAIL Skill packages the Secure AI Lifecycle catalog, 91 risks across the lifecycle, as a skill you can run gap assessments against.

claude skill add pillar-labs/sail-skill

Two warnings. It has 136 stars, the smallest project here, so the maintenance question is open. And its license field reads NOASSERTION, which means GitHub couldn't match it to a standard license. For a compliance artifact, unclear licensing is an odd place to start. Read the repo terms before you build a deliverable on it.

Security control mapping at volume

Anthropic Cybersecurity Skills is 754 structured skills mapped across MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, and the NIST AI RMF, sitting at 27,032 stars.

claude mcp add anthropic-cybersecurity-skills -- npx -y mukul975/Anthropic-Cybersecurity-Skills

Despite the name, this isn't published by Anthropic. It's a community project. The value for compliance work is the cross-framework mapping: when a NIST CSF subcategory needs to be traced to a technique and then to a control you already operate, that translation is exactly the kind of lookup a model does well and a person does slowly. It's a lot of context to load, so install it for mapping work rather than leaving it on permanently.

An order of operations

For a team heading into a first SOC 2 or ISO 27001:

1. Install the GRC pack. Run a gap assessment against the framework you're certifying to. Expect a list, not a verdict. 2. Install Prowler. Scan the cloud accounts in scope. Have the agent map findings to control IDs and mark the ones that don't apply, with a reason recorded for each. 3. Install SkillSpector. Scan every skill your team has installed, and keep the output. 4. Decide on SkillHub before your engineering team doubles. Retrofitting a registry is worse than starting with one. 5. Add SAIL only if you're certifying to an AI standard or shipping an AI product into a regulated market.

Steps 3 and 4 are the ones nobody does. They're also the two where an auditor's question currently has no good answer at most companies.

What none of this replaces

Evidence collection over time. Auditor relationships. Someone accountable for the risk register. Access reviews performed and signed. A model can draft a policy in four minutes that would have taken a consultant a week, and the policy is fine. The audit asks whether people followed it for a year, and no skill answers that.

Treat these as a way to stop paying for framework literacy. Buy that once, at zero cost, and spend the saved budget on the parts that need a human name attached.

Related reading: our security audit skills guide covers the scanning side in more depth, contract review skills handles the vendor-agreement half of third-party risk, and the cybersecurity tools list has the wider tooling picture.

FAQ

Can I use an agent skill to write my SOC 2 policies? Yes, and the output is usually better than a downloaded template because it's written against your actual architecture. Have a person who understands your environment review every control statement before it goes to an auditor. A policy that describes a control you don't operate is worse than no policy.

Will my auditor accept agent-generated evidence? Evidence is evidence regardless of what produced it. A Prowler scan output is the same file whether a human or an agent ran it. What auditors reject is generated narrative presented as observed fact. Keep the tool output and the commentary separate.

Are agent skills themselves a compliance problem? They're third-party code with local execution rights, so they fall under vendor management and change control at most organizations. Screening them with SkillSpector and serving them from a controlled registry like SkillHub is the cheapest way to have an answer ready.

Which of these covers the EU AI Act? The GRC pack includes EU AI Act, ISO 42001, and the NIST AI RMF as separate skills. SAIL complements it with a risk catalog specific to AI systems rather than a control framework. For an AI product in the EU, you want both.

Is the free GRC pack good enough, or do I need a compliance platform? Different jobs. The skills give you framework knowledge and drafting speed. Platforms like the commercial GRC vendors continuously collect evidence, which is the expensive part and the part skills don't touch. Small teams can run on skills plus spreadsheets much longer than vendors suggest. At audit scale, evidence automation earns its price.

How current are these frameworks? The GRC pack was last updated July 20, 2026 and Prowler on July 31, 2026. Standards move slowly, but regulatory deadlines don't. Check the repo's commit history against your certification date rather than assuming currency.