Security leaders are increasingly considering granting more authority to AI agents within Security Operations Centers (SOCs). According to a recent a16z survey of Chief Information Security Officers (CISOs) from companies with over 1,000 employees, 36% are currently comfortable allowing AI to autonomously manage low-risk security tasks. Within the next year, this figure is expected to rise to 60%, highlighting AI SOC as a significant initiative planned for 2027.
Driving Forces Behind AI Autonomy
The push towards AI-driven autonomy in SOCs is driven by mounting pressures. The same survey identified that 55% of respondents are concerned that AI-accelerated attacks could outpace their teams’ response capabilities. Analysts’ time, a critical asset in SOCs, is often consumed by gathering evidence from multiple systems to assess threats, making efficiency crucial.
Despite the willingness to embrace AI, security leaders remain cautious about full autonomy. At present, only 1% of respondents are comfortable with AI handling most security actions independently, with a slight increase to 7% anticipated in the next year. This indicates a desire for increased autonomy, but with careful management.
Understanding AI Autonomy Levels
The concept of AI autonomy isn’t binary; it’s a spectrum. Discussions often depict it as a slider from ‘AI suggests’ to ‘AI acts’, but this simplification can stall progress. Autonomy should be considered a series of decisions, specific to each workflow stage and action, where AI earns trust through evidence.
Historical frameworks, such as those by Parasuraman, Sheridan, and Wickens, offer insights into automation’s four broad functions: information acquisition, analysis, decision-making, and implementation. These can be applied to security operations, allowing different levels of automation based on the function and risk involved.
Implementing AI SOC Autonomy
Security teams need a common language to define autonomy levels for each workflow stage. Researchers at the University of Washington have proposed a five-level model based on human roles, from ‘Operator’ to ‘Observer’. This framework helps SOCs determine how much autonomy to grant AI agents.
The a16z survey suggests a trend towards moving from current L2 levels (human and agent collaborate) to L4 (agent works independently with oversight). Organizations should ensure that their AI SOC solutions can adapt both ‘human in the loop’ and ‘human on the loop’ models as needed.
Strategizing AI Autonomy in Security Operations
Deciding the level of autonomy for each SOC action involves evaluating reversibility, impact radius, and the strength of evidence. Actions like querying logs can be fully autonomous (L5), while more impactful decisions like isolating a production server might require a human decision (L2).
Security teams must establish criteria for expanding or reducing autonomy. This involves tracking AI accuracy, conducting test runs, and ensuring evidence trails for auditability. Autonomy should expand based on data, with readiness for demotion if necessary.
Ultimately, granting autonomy to AI SOC agents increases potential risk, necessitating safeguards similar to those for privileged identities. This includes assigning unique identities, restricting permissions, and logging all actions for human review.
As organizations move towards increased AI autonomy, they must balance the benefits with potential risks. By following structured approaches and maintaining robust oversight, security leaders can leverage AI to enhance efficiency while safeguarding their operations.
