OWASP Top 10 2026: Agents become the new attack surface engineering leaders

The 2026 OWASP LLM Top 10 reframes AI security around agentic risk, with excessive agency jumping to #3. The core shift: AI is no longer just generating content but acting autonomously on systems. Prompt injection remains #1 by practitioner concern, but incident data

shows it is effectively mitigated. The real surprise is misinformation rising in real-world incidents while practitioners underweight it. The panel emphasizes that agents must be treated as privileged identities with access controls, not just software. The core tension is functionality vs. security—agents need

broad access to be useful, creating a new attack surface where manipulation of intent, not just code, is the threat. The IBM team's key insight: controls that worked were those where AI never got a deciding vote. The project leads' blunt advice—"stop trying

to build a model that cannot be fooled; build the system so nothing important breaks when it is"—mirrors cyber resilience principles. The operational consequence: teams must shift from checklist compliance to tabletop exercises that test detection, containment, and reconstruction of AI-driven incidents.