Adversa AI Ranks 100 Agents by Security Vulnerability
TL;DR. Adversa AI evaluated 100 AI agents using its AI Risk Quadrant, finding that only 11 are both capable and well-defended against compromise. - A core problem is the 'lethal trifecta' of private data access, untrusted content exposure, and outbound actions common to 98% of tested agents. - The report highlights a 'power-protection inversion', where agents with greater capability often have wider attack surfaces. - Computer agents and coding agents show the most significant power-protection inversion among the categories reviewed.
- Adversa AI assessed the security performance of 100 AI agents.
- Only 11 of the 100 agents were deemed both capable and securely defended.
- The 'lethal trifecta' (private data access, untrusted content, outbound actions) is present in 98% of agents, hindering security.
- 'Power-protection inversion' means highly capable agents often have more attack surfaces.
- Computer and coding agents exhibit the highest power-protection inversion.
Sources
- Security of 100 AI Agents Tested and Ranked – What You Need to Know — securityweek.com