The AI security industry has stopped chasing model safety and started securing the harness.
OWASP’s Top 10 for Agentic Applications identifies attack vectors centered on harness control, not model behavior: goal hijacking, tool misuse, privilege abuse, code execution, memory poisoning, and cascading failures. The framework reflects broad consensus from 100+ contributors including NIST and Microsoft’s AI Red Team.
The bottleneck: a weak harness defeats a capable model. Agents make thousands of routing decisions. Every response can contain adversarial content. Prompt injection remains the top LLM risk with no reliable defense. The fix isn’t a smarter model; it’s deterministic policy engines enforcing permissions at execution time, outside the model.
Harness engineering is unglamorous: tool sandboxing, runtime policy checks, supply chain verification. This stops agents from pivoting between services or following injected goals. The model is no longer the constraint.
Sources: Distributing Security Controls Through Harness Engineering • OWASP Top 10 for Agentic Applications 2026
Coverage: The Harness Is the Hard Part of Agentic Pentesting
Related on this blog: PenClaw Sells a No-Refusals Pentester for $20 a Month. First You Show It Your Passport. • [AUTO] Data Diodes as AI Training Boundaries • [AUTO] BragJack: One Malicious Extension, Five Browsers