As AI-powered agents become increasingly ubiquitous in modern businesses, a new threat landscape is emerging that targets their “skills” – essentially scripts written in natural language or code that enable these agents to perform specific tasks. A recent cyberattack on an agentic AI work platform, Paperclip, highlights the risks posed by skills and has prompted the Open Worldwide Application Security Project (OWASP) to release a first-of-its-kind list of top 10 security issues for agentic skills.
The incident in question involved a malicious domain that impersonated Paperclip and distributed Trojanized Python packages and weaponized AI skills, compromising users’ machines and stealing sensitive information. The attackers were able to evade detection, with over 300,000 installs of the compromised skills within hours. This showcases why OWASP deemed it necessary to poll practitioners on this type of risk and create a standardized list of security issues for agentic skills.
OWASP’s new list highlights the dangers posed by skills and emphasizes that agent developers must address these risks to prevent malicious actors from exploiting them. According to Omar Turner, a security practitioner supporting the effort, many discussions at recent conferences like Black Hat USA focused on agents but neglected to consider the risks associated with skills. “Without the knowledge of what risks could exist with skills, you’re kind of operating blind,” he says.
Skills are essentially scripts that represent untrusted input in agentic architectures. They often rely on natural language and can reference external resources such as websites or model context protocol (MCP) servers. As Niv Hoffman, co-lead of the OWASP project, notes, skills are a major source of untrusted input and require secure-by-default environments to mitigate risks.
The OWASP Agentic Skills Top 10 list identifies two critical security risks: Malicious Skills and Supply Chain Compromise. The top risk is Malicious Skills, but the second critical risk is Supply Chain Compromise, which highlights the vulnerability of skills hosted on GitHub and other repositories. These skills can be used to pull typosquatted dependencies, allowing attackers to compromise agents quickly.
The OWASP effort aims to standardize a YAML format for skills files that automated analysis tools can use to determine legitimate from malicious skills. This is crucial for organizations to identify potential risks associated with agentic skills and develop secure-by-default environments. As Hoffman emphasizes, while it’s essential to allow people to adopt and use skills, “you must have them in a secured way” to unlock the full capabilities of agents.
In light of these findings, it’s crucial for organizations to reassess their approach to agentic security and prioritize securing skills from malicious actors. This includes implementing secure-by-default environments, conducting regular security audits, and staying up-to-date with OWASP guidelines on agentic skills risk management. By doing so, businesses can minimize the risks associated with AI-powered agents and prevent cyberattacks that exploit vulnerabilities in their skills.
Source: Dark Reading — 2026-08-21