Introducing Agent Skills Security: Identify and Prevent Risk in the AI Agent Supply Chain

Organizations are adopting Claude Code, Cursor, Gemini, Codex, and other agent ecosystems to accelerate development and automate increasingly complex tasks. As these agents become more capable, they also become more dependent on a new class of artifact: instructions.

Instructions, skills, workflows, configurations, hooks, and subagents increasingly determine how AI agents behave, what tools they use, what systems they access, and what actions they take.

Yet most security teams have little visibility into them.

Today, we're introducing Agent Skills Security, a new Heeler capability that helps organizations discover, analyze, and prioritize risk across the artifacts that drive AI agents.

A New Supply Chain Attack Surface

For years, software supply chain security has focused on two primary components:

• Source code
• Third-party dependencies

Security teams have built programs, tools, and processes around understanding and securing both.

AI agents introduce a third layer.

The instructions that guide agent behavior are increasingly becoming part of the software supply chain. These files can shape how agents make decisions, what actions they perform, what external systems they interact with, and what permissions they attempt to use.

In many organizations, these artifacts are already being committed to source control and shared across teams.

The challenge is that they largely exist outside the visibility of traditional security tooling.

Understanding the Risks

Agent artifacts can introduce risks that are fundamentally different from traditional application security issues.

Security teams need to understand questions such as:

• Is an agent being instructed to perform unsafe actions?
• Does a skill introduce risky or unexpected behavior?
• Are external systems being referenced that warrant investigation?
• Could an instruction influence an agent to bypass established controls?
• Are developers unknowingly introducing risky agent capabilities into repositories?

Without visibility into these artifacts, organizations risk creating blind spots as AI adoption accelerates.

Introducing Agent Skills Analysis

With Heeler, organizations can now:

• Discover agent artifacts across repositories
• Inventory instructions, skills, workflows, and configurations
• Analyze files for suspicious or malicious behavior
• Identify risky agent artifacts that warrant review
• Understand why a file is considered risky through transparent findings and scoring
• Prioritize remediation efforts based on risk

Rather than treating these files as isolated assets, Heeler incorporates them into a broader understanding of the software supply chain.

Combining Deterministic and Semantic Analysis

Agent Files analyzes artifacts using multiple approaches.

Deterministic analysis identifies structured signals such as suspicious external references and other risky patterns.

Semantic analysis evaluates the intent and behavior described within the artifact itself, helping identify risks that traditional pattern matching alone may miss.

The result is a more comprehensive understanding of both what an agent artifact contains and what it is attempting to accomplish.