Why third-party AI skills are the next blind spot, even on enterprise AI
Every few weeks, a new file does the rounds among AI users. Someone shares a pre-built “skill”, a plugin or a short Markdown file that promises to make your AI assistant faster, smarter or cheaper to run. Install this, and it will write better reports. Add that, and you’ll use fewer tokens (the units AI usage is billed in). It takes seconds, and because the business pays for an enterprise licence of a leading AI model, it feels safe, but it isn’t necessarily.
What you are actually installing
A skill is not just a set of tips. It is a package of instructions, sometimes with code attached, that tells an AI agent how to behave and which tools to use. Once installed, it can run with the same access as the agent itself: your files, your email, your connected systems and, in some setups, the credentials stored on your machine.
That means a single downloaded file can quietly instruct your AI to send data somewhere it shouldn’t, or to change how it handles a task, without anyone noticing.
Primary Risks
- Prompt injection – Malicious Instructions hidden inside a skill file or bundled script can trick AI assistant into executing unauthorized actions.
- Data & credential Exfiltration – A Compromised skill can scan you machine for sensitive files, API keys or cloud tokens and send them to an external server.
- Arbitrary code execution – Skills often instruct claude or your system to pull down external packages, run back commands, or pipe URLs directly into the terminal
This is already happening
In September 2025, researchers at Koi Security found what they described as the first malicious MCP server seen in the wild. MCP (Model Context Protocol) servers are a close cousin of skills: small add-ons that connect an AI assistant to other tools, and they raise the same trust problem. This one was an impostor copy of Postmark’s email integration, published on npm, a public library of code packages; Postmark itself had no involvement. It worked exactly as expected for fifteen versions. Then version 16 added a single line that blind-copied (BCC’d) every email sent through it to an outside address. According to The Hacker News, it had been downloaded more than 1,600 times.
In February 2026, security firm Snyk scanned almost 4,000 skills from public skill marketplaces for its ToxicSkills research. More than a third had at least one security flaw, and 76 contained confirmed malicious payloads, designed to steal credentials, install backdoors or switch off the agent’s safety checks. Publishing a skill on some of these marketplaces required little more than a Markdown file (a simple text document) and a GitHub account just one week old, according to Snyk.
In the same month, Koi Security reported 341 malicious skills on ClawHub, a public skill marketplace, 335 of them from a single campaign it called ClawHavoc. Disguised as everyday tools, including finance utilities and workplace integrations, they tricked users into installing software that stole passwords and other credentials.
Why an enterprise licence doesn’t cover this
An enterprise subscription protects how the AI provider handles your data: retention, training and security on their side. Unless administrators have locked down what can be installed, it doesn’t vet what your people choose to plug into it. A skill downloaded from a forum and added by an employee runs with that employee’s access. To your systems, its actions look like theirs. In practice, the biggest exposure is where those controls are loose: staff trying out skills on personal or developer setups, or on platforms where installation hasn’t been restricted.
Picture a finance team that uses an AI assistant to prepare payment instructions and supplier onboarding documents. A poisoned skill quietly changes one supplier’s bank account number in a draft. Unless someone checks the details against the supplier master data, the payment looks routine and the money goes to a fraudster’s account.
That is what makes this a fraud risk, not just an IT one. A compromised skill can:
- harvest passwords and access keys, which then open the door to payment diversion or business email compromise
- copy client data, contracts or price-sensitive information to an outside server
- alter what the AI produces, such as a figure in a reconciliation or a management report, in ways a busy reviewer could miss
- influence decisions by feeding the AI skewed instructions or data
- lie dormant for weeks before activating
Why it is so hard to detect
Traditional fraud leaves a trail of unusual people, unusual payments or unusual access. Here, the activity comes from a trusted tool, under a legitimate user’s name, inside an approved licence. Skills are often plain text, so they don’t look like software to the people installing them, or to some security tools.
When something does go wrong, tracing it back means working out which skill was installed, by whom, when, what it could reach and what it did. That can be slow and difficult, and so can making sure every trace has been removed.
Five questions to ask before anyone installs a skill
- Where did it come from? Is the publisher known and accountable, or is it an anonymous upload?
- What does it actually contain? Has someone with the right expertise reviewed its instructions and any attached code?
- What can it reach? Does it need email, file or system access to do its job, or is it asking for more?
- Who approved it? Is there a list of approved skills, and is installation restricted to it?
- Would we know if it misbehaved? Are the agent’s actions logged and reviewed?
Whatever produced a payment instruction, a person or an AI, keep the control your fraud team already trusts: verify any change to bank details independently, through a known contact and a separate channel, before money moves.
If someone has already installed a skill they can’t vouch for, they should tell IT or security and stop using it until it has been reviewed.
If the answer to any of these questions is “we don’t know”, the convenience isn’t worth the risk.
“Skills are one of the most useful things to happen to enterprise AI. They let you turn a general model into something that knows how your business works. But a skill is code that acts with your access, and it deserves the same scrutiny as any software you’d let near your finance systems. Build them, or vet them, before you trust them.”
Shubham Bindal, Director – AI & Digital, Pierag Consulting
How Pierag helps
Pierag’s AI & Digital team designs and builds custom AI skills, agents and automations for finance and business teams, developed in-house, reviewed before release and built around least-privilege access and audit trails. We apply the same standards to the tools we use ourselves.
This Fraud Awareness Week, the question isn’t whether your teams are using AI. It’s whether you know what they’ve plugged into it. [Talk to our AI & Digital team]
Sources:
- The Hacker News: First malicious MCP server found
- The Register: Postmark MCP server code hijacked
- Snyk: ToxicSkills, malicious AI agent skills on ClawHub
- Snyk Research: Agent Skills threat landscape
- Koi Security: ClawHavoc, 341 malicious skills on ClawHub