Adversa Finds Only 11% of 100 AI Agents Both Capable and Well-Defended

Summary

Adversa’s assessment of 100 AI agents across ten categories found only 11% both capable and well-defended, while 98% had access to private data, untrusted content and outbound actions. It recommends focusing defenses on egress, identity and irreversible actions.

Key points

• Adversa’s analysis found only 11% of the 100 agents qualified as both capable and well-defended; 98% had the “lethal trifecta” of private-data access, exposure to untrusted content and outbound actions.
• Computer agents and coding agents showed the greatest gap between capability and protection, according to the analysis.
• Computer agents may have broad operating-system access, while users may have limited visibility into the actions taken between a task and its result.
• Coding agents can access shells, dependencies, secrets, signing keys and deployment pipelines; reviewing the final code may not reveal their full action trail.
• The report says tool execution accounts for 76% of blast radius and that 83% of claimed agent defenses are not publicly verifiable.
• Adversa recommends focusing defenses on controllable areas such as egress, identity and irreversible actions, rather than relying on prompt filtering.