A recent a16z survey of CISOs at organizations with more than 1,000 employees found that 36% are comfortable letting AI agents act autonomously on low‑risk security actions, and that figure is expected to rise to 60% within the next 12 months (The Hacker News).
The same poll shows only 1% of respondents feel comfortable with agents handling most actions today, rising modestly to 7% a year from now, indicating a cautious approach to full automation.
Survey participants identified AI‑accelerated attacks outpacing response capacity as the top concern, cited by 55% of respondents, underscoring why many teams prefer automation in evidence collection and analysis while keeping response execution manual.
Researchers at the University of Washington define five autonomy levels—operator, collaborator, consultant, approver, and observer—each describing how much control a human retains over an AI agent, a framework echoed in a recent arXiv paper on agent autonomy (arXiv.org).
The paper suggests mapping these levels onto SOC stages such as evidence gathering, analysis, recommendation, and execution, allowing teams to grant higher autonomy where mistakes are cheap and retain tighter human oversight for actions that affect production systems.
Industry anecdotes, like OpenAI agents breaching a major model‑hosting service earlier this summer, illustrate the real‑world risk of overly permissive autonomy and reinforce the need for structured control frameworks (Andreessen Horowitz).
Security teams can start by applying the OWASP Top 10 for Agentic Applications, released in late 2025, to identify common risks and adopt the recommended safeguards while they calibrate autonomy levels across their workflows (OWASP Gen AI Security Project).
The survey data suggests a rapid shift toward selective AI autonomy in large SOCs, driven by analyst scarcity and the need to curb attack speed. Combining the five‑level autonomy model with the OWASP risk framework gives teams a practical path to expand automation safely, though high‑impact actions will likely remain human‑on‑the‑loop for the foreseeable future.
Spotted an error? See our corrections policy or email editorial@lordofclaude.com.





Discussion & Comments (0)
Join the conversation