OWASP establishes security framework for agentic AI

The OWASP Top 10 for Agentic Applications has established a security framework to address vulnerabilities in autonomous AI agents. As these systems integrate into professional workflows, they introduce new attack surfaces through their “skills”—a combination of text instructions and code.

Research indicates that over 35% of skills analyzed on public marketplaces contain security vulnerabilities. For instance, the ClawHavoc campaign successfully infected more than 1,100 skills on the ClawHub platform, enabling attackers to steal credentials and cryptocurrency wallets. Traditional code analysis tools often fail to detect these threats because attacks can be hidden within natural language instructions, Markdown files, or seemingly legitimate text descriptions.

To combat these risks, OWASP proposes a universal YAML format to standardize permission declarations, content hashing, and access control across autonomous agent platforms. The current state of agentic AI is being compared to the early internet era, where rapid capability development has outpaced the industry's understanding of security implications regarding persistent memory, API access, and autonomous decision-making based on untrusted content.