OWASP Agentic Skills Top 10
Analytics2026-09-17, 10:05
OWASP has released the first complete version of Agentic Skills Top 10. This is a separate Top 10 focused on risks associated with AI agent skills.
The list is fairly "classic". Much of it is already familiar from traditional AppSec, although agents do introduce some specific considerations of their own.
The Top 10 are:
- Malicious Skills - skills that are malicious by design
- Supply Chain Compromise - compromise of the author, registry, or dependencies
- Over-Privileged Skills - excessive permissions
- Insecure Metadata - issues in manifests, YAML and metadata
- Untrusted External Instructions - the skill pulls unsafe instructions from external sources
- Weak Isolation - execution without proper sandboxing or isolation
- Update Drift - the skill's behavior changes after an update, and the user may not even notice
- Poor Scanning - scanners detect known patterns but miss attacks that operate at the semantic, contextual, and behavioral levels
- No Governance - no proper inventory of skills, change control, approval process, or audit logs
- Cross-Platform Reuse - a skill is moved between platforms, and some of its security context and restrictions are lost along the way.
OWASP cites a Snyk study in which 36.8% of 3 984 reviewed skills contained security weaknesses, and 13.4% were classified as critical. Confirmed malicious payloads were also found in more than 70 skills.
The document also includes examples of supply chain attacks, dependency hijacking, and skills that bypassed existing security scanners.
The document itself is more interesting than just a list of ten items. In the Executive Risk Map, the risks are organized into four areas: Source & Registry Trust, Execution Boundary, Lifecycle Governance and Cross-Platform Reuse. Each risk is then analyzed using the same template: description, skill-specific considerations, real-world cases, attack scenarios, safeguards, mappings to other OWASP/CSA frameworks, and related risks. In practice, this provides a ready-made threat modeling framework for AI skills.
Vendors
Products