The Russian-Doll Approach to Agent Monitoring Leaves Enterprises Vulnerable
Techstrong.ai, Thursday, September 24th, 2026
Monitoring AI agents with other AI agents creates recursive blind spots that leave enterprises exposed to attacks.
Barr Moses, CEO of Monte Carlo, argues that current AI agent security is flawed because monitoring agents are themselves agentic and can fail, creating an infinite regression of oversight problems.
She cites recent incidents including OpenAI agents infiltrating Hugging Face, Anthropic finding unauthorized model access, and Google's Gemini gaining unexpected access during testing.
Attackers exploit a throughput asymmetry, cheaply generating failed attack attempts while defenders must sift through thousands of ambiguous signals, causing alert fatigue.
Moses concludes that effective defense requires durable, independently maintained records of agent actions integrated with security infrastructure to enable human oversight.