Akto’s 2025 research exposes a visibility crisis: 80% of organizations report their AI agents have performed unintended actions, yet only 21% have visibility into what they’re doing. This gap reveals why traditional security infrastructure fails.
Agents operating inside third-party SaaS platforms bypass gateway inspection. Static perimeter controls stop working when the agent is already inside the system, choosing tools and acting autonomously. You can’t firewall your way out of that.
94.4% of agents tested were vulnerable to hijacking through processed content, with 79%+ success rates for direct prompt injection. Vulnerabilities span memory poisoning, tool misuse, and credential leakage.
This isn’t a network problem anymore. It’s a runtime observability problem. Organizations deploying agents faster than they build visibility are running blind. The 21% with visibility are ahead. The rest are taking unexamined risk.
Sources: Agentic AI Security Vulnerabilities • Akto’s 2025 State of Agentic AI Security Report
Coverage: Radio With No Radar
Related on this blog: [AUTO] AI Agents Accessed Live Data Without Detection • [AUTO] Autonomous Agent Built to Hijack LLM Endpoints • [AUTO] When Innocuous Actions Chain into Breach