Cybersecurity Skills Review 2026: AI Security Training

Cybersecurity Skills is not a training platform in the usual sense. It is a repository of 817 structured security skills for AI agents, mapped to MITRE ATT&CK, NIST CSF 2.0, D3FEND, and three more frameworks. The GitHub repository at github.com/mukul975/Anthropic-Cybersecurity-Skills shows 34,564 stars. The website is mahipal.engineer/Anthropic-Cybersecurity-Skills/. Pricing starts at $0/mo. That is the whole product surface. There is no video library, no lab sandbox, no certification voucher, no seat license. So the honest question for a 2026 buyer is not "is this a good course" but "what is this, and does it belong in a security training budget at all."

This review works from the repository, the vendor documentation, the pricing page, and live job-posting data. It does not claim lab completion timestamps, sandbox costs, or a threat-model run, because those were not performed. Where the brief asked for hands-on evidence, this review states plainly what the public material supports and what it does not.

What Cybersecurity Skills covers

The repository describes a skill pack: 817 discrete security skills written for AI agents to execute. Each skill is a structured unit, not a lesson. The mapping to MITRE ATT&CK, NIST CSF 2.0, D3FEND, and three additional frameworks is the core design decision. It means a skill is indexed against a known adversary technique or control function rather than against a chapter number.

That is a different shape from a course catalog. A course catalog answers "what will I learn this week." A skill pack answers "what can an agent be told to do, and under which framework does that action sit." The repository does not present itself as certification prep. It does not present itself as a video course. It does not present itself as a hands-on lab environment. It presents itself as a structured capability library.

The scope is broad by framework coverage and narrow by delivery. Broad because ATT&CK, CSF 2.0, and D3FEND together span offensive technique, defensive control, and countermeasure mapping. Narrow because there is no runtime. The skills are definitions. Something else has to execute them.

Version and syllabus dating matter here. The repository is the source of truth for what exists at any moment, and it changes. A buyer evaluating this in 2026 should read the repository at the commit they intend to use, not a blog summary of it. The website is a landing surface; the repository is the artifact.

Pricing 2026: individual vs team

The pricing page lists a starting tier of $0/mo. That is the only figure the public material supports. There is no published seat cost above that, no lab-hour meter, no certification voucher bundle, and no team tier with a stated per-seat number in the material reviewed.

This is worth pausing on, because it changes the buying conversation. A conventional cybersecurity training platform in 2026 sells seats, meters lab hours, and bundles exam vouchers. Those three line items are how training budgets get justified. Cybersecurity Skills has none of them in the public pricing. The $0/mo starting tier means the cost model is not per-learner consumption. It is closer to an open artifact that a team adopts.

For an individual, that removes the usual friction. There is no trial window to evaluate, no per-seat negotiation, no lab-hour overage to watch. For a team, it removes the usual procurement path. There is no seat count to buy, so there is no seat count to justify. A security team adopting this is really making an engineering decision about whether to build agent workflows on top of the skill definitions, not a training decision about how many people get licenses.

I would not treat the $0/mo tier as a free trial of a paid product. The material does not describe a paid product above it. Treat it as the product.

Curriculum audit: hands-on vs video

There is no video. The ratio of lab time to lecture time is not a meaningful measurement here because neither exists in the conventional sense. The unit of content is the skill, and the repository shows 817 of them.

What replaces the lab-to-lecture ratio is a coverage ratio: how many of the 817 skills map to each framework, and how deep the mapping goes. The repository states the frameworks. It does not, in the material reviewed, publish a per-module exercise count in the way a course would. So the audit has to be structural rather than quantitative.

Structurally, the design is closer to a control library than a curriculum. Each skill is a capability an agent can invoke. That means the "exercise" is whatever the adopting team builds around the skill. The repository supplies the definition; the team supplies the scenario. This is a real difference from a platform that ships pre-built exercises with expected answers.

For a learner who wants guided progression, that is a gap. For an engineer who wants a machine-readable security capability set to wire into an agent, it is the point. The same artifact serves one audience and frustrates the other, and the repository does not pretend otherwise.

Hands-on test: three labs start to finish

This section would normally report setup friction, broken steps, and sandbox cost across an entry, mid, and advanced lab. That test was not run, and inventing it would be worse than omitting it. What can be said is what the public material supports about the execution model.

There is no sandbox in the product. The repository does not ship a hosted environment where a learner clicks into a vulnerable VM. Execution depends on the agent runtime the adopting team supplies. That means setup friction is not a property of Cybersecurity Skills; it is a property of whatever stack the team points at it. Sandbox cost is likewise external. If a team runs skills against a cloud account, the cloud bill is theirs, not the vendor's.

This is the single biggest structural difference from a hands-on lab platform. A lab platform owns the environment, so it owns the friction and the cost, and it can promise a known-good path. A skill pack owns neither. The adopting team owns both. For a team with an existing agent runtime and cloud security tooling, that is fine. For a learner with no runtime, there is nothing to click.

The practical implication: evaluate this the way you would evaluate a library, not a course. Read the skill definitions, check whether the framework mappings match your control framework, and decide whether your runtime can execute them. That is the real acceptance test, and it is one the buyer runs, not the vendor.

AI angle: does it teach securing AI systems

This is where the product is most interesting and most easily misread. Cybersecurity Skills is a security capability set for AI agents. It is not, on the evidence of the repository description, a curriculum about securing AI systems. Those are different things, and conflating them is the most common mistake a buyer will make in 2026.

Securing AI systems covers prompt injection, model supply chain integrity, agent permission boundaries, and data exfiltration paths. A training platform that teaches those topics would have modules on each. The repository description does not claim that. It claims 817 security skills for AI agents, mapped to security frameworks. The agent is the executor, not the subject.

That distinction matters for anyone shopping for an AI security certification. If the goal is to learn how to defend a model deployment, this is not that artifact. If the goal is to give an agent a structured set of security actions it can take, this is exactly that artifact, and the framework mapping is what makes it auditable.

There is a second-order benefit worth naming. A team that adopts this is, by necessity, thinking hard about what an agent is allowed to do. Wiring 817 security skills into an agent forces the permission question: which skills can this agent invoke, under what conditions, with what logging. That is adjacent to the agent-permissions problem even though the repository is not a course on it. The work of adoption surfaces the question. The repository does not answer it for you.

For teams already running agents against security tooling, the framework mapping is the useful part. ATT&CK and D3FEND give a shared vocabulary for what a skill does, which is what you need when an agent takes an action and someone asks why. NIST CSF 2.0 gives the governance framing. The three additional frameworks extend that coverage. None of this is AI security education. All of it is AI security infrastructure.

Cert and job value: what employers ask for

Cybersecurity Skills does not issue a certification. The repository does not describe an exam, a badge, or a credential. So the job-market question has to be answered honestly: this product does not move the certification line on a resume.

What employers ask for in security roles is a separate question from what this product provides. Live job postings in 2026 continue to name established certifications and framework familiarity. The framework familiarity part is where this product has indirect value. A candidate who can speak fluently to ATT&CK technique mapping, CSF 2.0 function alignment, and D3FEND countermeasure selection is demonstrating exactly the vocabulary that security job postings ask for. The repository is a way to build that fluency, but it is not a credential that proves it.

So the match rate against live postings is effectively zero on the certification axis and non-trivial on the framework-vocabulary axis. A hiring manager scanning for a named cert will not find one here. A hiring manager asking a candidate to explain how a control maps to a framework will get a better answer from someone who has worked through this material than from someone who watched a video series.

I would not put this on a resume as a certification. I would put the framework fluency it builds into an interview answer, and I would let the work speak. That is a smaller claim than a cert, and it is the accurate one.

Limits and who should skip

The prerequisite depth is the first limit. This is not an entry point into security. A reader who does not already know what ATT&CK is, or why CSF 2.0 restructured around functions, will find the skill definitions opaque. The repository assumes framework literacy. It does not teach it.

The second limit is the runtime dependency. Without an agent runtime, there is nothing to execute. A learner looking for a guided path with a known-good environment will not find one here. That is not a defect in the artifact; it is a mismatch with the audience.

The third limit is maintenance. A skill pack mapped to frameworks has to track framework revisions. ATT&CK and CSF 2.0 both evolve. The repository is the live artifact, and its currency depends on ongoing maintenance. A team adopting it should plan to track upstream changes rather than treat it as a frozen reference.

Who should skip: anyone shopping for a certification, anyone who wants video instruction, anyone who needs a hosted lab, and anyone who wants a course on securing AI models rather than a capability set for AI agents. That is a large group, and it is worth being blunt about it. The product is narrow on purpose.

Verdict

Cybersecurity Skills is a well-scoped artifact that is easy to misbuy. The 817 skills, the framework mappings, and the 34,564-star repository are real and specific. The $0/mo starting tier is real. What is not real is the expectation that this is a training platform with labs, videos, and a certification at the end. It is a machine-readable security capability set for agents, and it should be evaluated as infrastructure.

Score: 7.5/10. High marks for framework coverage and clarity of purpose, lower marks for the absence of any guided path, certification, or hosted environment, which limits who can actually use it.

Use this if you run AI agents against security tooling and want a structured, framework-mapped set of security skills to give them, or you want to build fluency in ATT&CK, CSF 2.0, and D3FEND vocabulary.

Skip if you want a certification, a video course, a hosted lab, or a curriculum on securing AI models.

Beetlix is our own product. Where Cybersecurity Skills supplies framework-mapped skill definitions for agents, Beetlix is a different kind of tool, and the two are not substitutes. If you are evaluating agent infrastructure more broadly, it is worth looking at both, but do not buy one expecting the other.

How this review was researched

Sources: the vendor documentation and website at mahipal.engineer/Anthropic-Cybersecurity-Skills/, the official pricing page, the GitHub repository at github.com/mukul975/Anthropic-Cybersecurity-Skills (34,564 stars), and live AI model pricing data. No lab runs, sandbox tests, or completion timestamps were performed for this review, and none are reported.