AI Engineer World's Fair 2026
We Vetted 2,000 AI Skills Before They Reached Developers
Loading slides…
Read the talk
We Vetted 2,000 AI Skills Before They Reached Developers
Nubank’s skill review pipeline treats shared agent instructions as supply chain dependencies, combining local checks, mandatory CI, contextual review and execution controls.
From a talk by Lucas Palma
What enters the workflow before code is generated?
A developer adopts an AI skill to make coding faster. It looks like configuration, but it can change what the agent does before anyone reviews the resulting code. Where should security review begin? Lucas Palma, introducing himself as Product Security Manager at Nubank, describes a system for reviewing skills before developers receive them. His team’s remit includes helping engineers and product managers make products safer; shared AI instructions extend that responsibility into the development workflow itself.
Skills belong in the software supply chain because they influence behavior. Reviewing generated code alone leaves the instructions that shaped it outside the review boundary. The familiar dependencies—packages, containers and models—remain, but skills, plugins, MCP servers and agent rules now sit alongside them. The security boundary therefore has to cover the whole workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a helpful instruction to a credential leak
The credential example follows a short chain. A skill tells the agent to retrieve a token. The agent starts using the token as a hard-coded value, and that value can reach logs, creating a later data leak. The problem begins with an instruction about how to accomplish a task, before it becomes a problem in generated code or operational output.
Other risks depend on where the shared instructions run:
- Shell commands: An author’s command instructions are reused in another developer’s shell, where they can modify files or perform dangerous operations.
- Excessive permissions: A skill can receive more authority than its task needs, increasing the consequences of a mistake.
- Accidental errors: Harm does not require malicious intent. Even a typo can become dangerous in the wrong execution context.
Review therefore needs to examine both the instruction and the environment in which it could take effect.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make distribution the review boundary
Nubank’s internal plugin marketplace provides a canonical place to discover and share skills. Authors contribute through pull requests, and the plugins distributed through the marketplace can include skills. That creates a practical intervention point: review a contribution before it becomes available to everyone else. Palma calls the CI review tool Skill Vector. It assesses the submitted skill, classifies risks and requests remediation before marketplace distribution.
The pipeline supports fast author feedback without depending on authors to enforce the policy themselves:
- Scan locally. Developers can check a new or changed skill and iterate before submitting it.
- Rescan in CI. Submission triggers another scan because the system cannot assume that the author ran the local check or used an up-to-date scanner.
- Check known patterns deterministically. Regular expressions and other straightforward checks identify risks that do not require much interpretation.
- Review context with an LLM. Contextual analysis follows where understanding the instructions or behavior requires more than pattern matching.
The two analysis methods serve different purposes. Palma points to variability in LLM judgments, including the effect of temperature, as a reason to retain deterministic checks rather than hand the entire decision to a model.
Findings return to the originating pull request, so the author can see what needs to change where they are already working. The pipeline also emits SARIF, allowing security tools to consume the results and bring them into the vulnerability management program. Developer feedback and security tracking become two destinations for the same review findings.
Severity and policy determine what happens next: a skill may be allowed, require remediation or be blocked before distribution. The complete path runs from local scanning to the pull request, deterministic analysis, LLM review, PR feedback and SARIF, then the policy decision. Local checks shorten the feedback loop; CI makes the review part of publication.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Inspect behavior, authority and data handling
Skill Vector’s review surface goes beyond finding suspicious command strings. Palma gives a non-exhaustive set of targets:
| Review surface | What the scanner examines |
|---|---|
| Instructions and behavior | Unsafe guidance and drift in agent behavior |
| Shell and filesystem | Destructive commands and inappropriate file modifications |
| Credentials and data | Credential requests and unintended exposure |
| Permissions and integrations | Overbroad authority and risky MCP usage |
The credential question, for example, is not simply whether a skill mentions a token. It is how the skill asks for credentials and what happens to the data afterward. Similarly, a file operation needs to be assessed as an action in context, not merely as text appearing in a skill.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Count findings separately from affected skills
Palma reports a presentation baseline of over 2,000 skills scanned, noting that the total had already grown. He reports more than 1,500 risk findings, not more than 1,500 affected skills: a single skill can contain several risks. He says 1,000 of those risks were probably remediated shortly afterward. The qualification matters: these are reported operational counts, not a measured detection accuracy or a confirmed remediation rate.
Some especially risky submissions were blocked before they reached the marketplace. The rollout also included a historical scan of skills created before Skill Vector existed. Those older findings entered the vulnerability management program for remediation, extending the effort beyond controlling new submissions to addressing the existing catalog.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make findings useful in their actual context
The useful combination was deterministic scanning for known patterns, LLM review for behavior, checks around destructive commands and credentials, and delivery through SARIF and PR comments. The harder work was deciding which signals deserved attention. Comments initially received similar risk treatment even though their risk depended on context, and weak signals without enough context caused more trouble than they helped resolve.
Prompt instructions asking for confirmation also needed closer scrutiny. So did warnings that seemed harmless when a skill ran locally but could become consequential in production. A finding needed clear remediation guidance to be useful to the author. Coverage posed another operational problem: teams could create additional marketplaces. The security process therefore needed to discover those distribution channels proactively and bring Skill Vector into them too.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Put human approval at the execution boundary
An instruction to ask for confirmation does not establish human approval. Palma describes the failure mode: an AI may ask itself for confirmation, or accept another AI’s confirmation, and treat the requirement as satisfied. The person reading the skill believes a human is in the loop, while the agent sees a completed confirmation step.
A minimal skill instruction illustrates the ambiguity:
markdown
Before modifying files, ask for confirmation.
The sentence requests a behavior, but does not itself identify a human approver or enforce a pause in execution. The review needs to follow the action to the tool that will execute it: does that tool actually pass through a human approval gate? Palma points to approval gates and hooks as controls around execution. The relevant distinction is between asking the agent to obtain approval and requiring approval before an action can run.
Skills are only one component of a plugin marketplace, so the surrounding controls can be reused: review contributions before publication, let developers run checks locally, enforce those checks again in CI, and combine deterministic analysis with contextual LLM review. For dangerous actions, that review must connect to enforcement when the action occurs. Prompt wording alone cannot supply the execution boundary.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bring third-party components through the same gate
The broader scope includes plugins, MCP servers, rules and hooks, with different risks for each. Palma names MCP Vector and rules checks as related controls already in use, while describing further work on gates that vary with policy and marketplace context. The direction also includes tool-level enforcement, audit logs and trusted gateways, extending control from what enters the catalog to what happens during execution.
A trusted marketplace supplies the canonical route for scanning and sharing, including components written outside the organization. Downloading a third-party skill or plugin should not bypass the review process: it should enter the marketplace for scanning before use and wider distribution. That makes the adoption path part of the security design. A component is reviewed because it will influence the workflow, regardless of who wrote it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
Defines skill directories, instruction files, executable scripts, supporting resources, and optional tool declarations.
Explains how a security platform consumes SARIF results and which fields and format constraints integrations must satisfy.
Documents execution hooks and permission decisions, including examples that block destructive shell commands before execution.
Full transcript
Loading transcript…