Most Powerful AI Agents Lack Basic Defences, Study Finds

Adversa AI's analysis of 100 AI agents reveals concerning security issues, where only 11 agents are deemed 'capable well-defended.' A ‘lethal trifecta’ of private data access, exposure to untrusted content, and outbound action abilities poses significant risks, compromising both capability and security. The analysis highlights that most AI agents possess this trifecta, and the most powerful agents also tend to have the least protection.

Categories such as computer and coding agents exhibit the largest security gaps, as users lack visibility and control over their actions. Adversa recommends focusing on controlling outputs and managing data to mitigate risks, as current AI solutions are seen as black boxes with limited user input control.