OWASP Top 10 for Agentic Applications 2026
OWASP Gen AI Security Project · v2026 · first published 2025-12-09
Scope
The ten most critical security risks facing autonomous and agentic AI applications. Globally peer-reviewed, developed with 100+ industry experts and researchers. The most visible formal taxonomy of agentic AI risk.
Why it's ingested
OWASP Agentic Top 10 is the single most visible formal taxonomy of agentic AI risk. Ingesting it as a first-class Taxonomy node lets SAICA-KG crosswalk OWASP categories to its FailureMode facets and answer OWASP-keyed queries through the website and JSON API. This is the "OWASP/ATT&CK for AI coding agent supervision" positioning operationalized.
Categories & crosswalks
Each row lists one external category and any SAICA-KG FailureModes that cross-walk into it.
ASI01Agent Goal HijackAttacker-induced diversion of an agent from its user-assigned goal.
SAICA-KG crosswalks:ASI02Tool Misuse and ExploitationAbuse of the agent's tool surface to exceed intended authority.
SAICA-KG crosswalks:ASI03Identity and Privilege AbuseMisuse of agent identities or escalation of privileges.
SAICA-KG crosswalks:ASI04Agentic Supply Chain VulnerabilitiesCompromise via the agent's software, tool, model, or data supply chain, including hallucinated-package-name typosquatting (slopsquatting).
SAICA-KG crosswalks:ASI05Unexpected Code Execution (RCE)Agent emits or runs code producing unintended remote execution.
SAICA-KG crosswalks:ASI06Memory & Context PoisoningAttacker manipulation of agent memory or long-context state.
SAICA-KG crosswalks:ASI07Insecure Inter-Agent CommunicationWeakly authenticated or validated messages between agents.
SAICA-KG crosswalks:ASI08Cascading FailuresErrors that propagate across agent subsystems or the agent fleet.
SAICA-KG crosswalks:ASI09Human-Agent Trust ExploitationSocial-engineering of humans by or via an agent.
ASI10Rogue AgentsUnsanctioned or self-modifying agents operating outside policy.
SAICA-KG crosswalks:
Notes
Full ASI descriptions will be expanded during v0.1 seeding; the category labels here are canonical from the OWASP Gen AI Security Project.