The Path to AI Autonomy in the SOC
Security leaders are about to hand AI agents a lot more authority. In a recent a16z survey of CISOs at organizations with more than 1,000 employees, 36% said they are comfortable today letting AI agents act autonomously on low-risk security actions. In 12 months from now, 60% expect to be comfortable at that level. Overall, AI SOC was the highest ranking initiative for 2027. The pressure
Security professionals are poised to grant AI agents greater decision-making power. A recent a16z survey of CISOs managing organizations with over a thousand employees indicates 36% currently feel at ease letting AI agents handle low-risk security tasks autonomously, with this comfort level expected to rise to 60% within the next year.
Overall, AI-driven security operations (SOC) is the top priority for 2027. The driving force behind this shift is the concern that AI will outpace human response times to attacks, cited by 55% of respondents. Analyst time remains the most precious resource in SOC investigations, as each alert necessitates evidence gathering from multiple systems, analysis, and judgment on the alert's significance.
However, when it comes to granting AI agents full autonomy, only 1% of respondents are comfortable with this today, and only 7% anticipate this comfort level in a year. Many teams desire more autonomy but are hesitant to relinquish complete control. Autonomy is often viewed as a simple scale ranging from AI suggesting actions to AI fully implementing them, but this oversimplification is why many programs fail.
Autonomy should be viewed as a series of decisions made at each stage of the workflow and for each action, earned through evidence rather than granted outright. Security leaders should focus on four stages of automation: information acquisition, information analysis, decision and action selection, and action implementation. By breaking down autonomy by stage, teams can more effectively allocate autonomy, automating the first two stages heavily where analyst hours are most critical and mistakes are cheapest to rectify, while keeping the final stage conservative where business disruption from errors is highest.
To facilitate this, a framework from the University of Washington defines five levels of agent autonomy based on the role of human oversight. Researchers have long applied a framework outlining four functions of automation: information acquisition, analysis, decision and action selection, and implementation. Each function can exist at varying levels of automation within the same system.
Applying this to security operations, teams can decide how much autonomy each stage should have. A rubric with three key questions can help teams determine the appropriate level of autonomy for each action. These questions ensure that each action is evaluated individually rather than relying on a single policy that may be too permissive or overly restrictive.
While 52% of CISOs prefer agents to require human approval for all actions, this approach may conceal the failure mode of inadequate context for reviewers. Proper approval steps should provide three key pieces of information: the AI's logic, evidence, and forensic data to support its decisions. This approach, known as "least agency," ensures agents only receive the minimum autonomy required for safe, bounded tasks, with autonomy expanding as it is earned.
To avoid risking too much, agents should operate at lower levels of autonomy initially, with criteria in place to promote autonomy upgrades. However, teams must also be vigilant about lowering autonomy if agreement rates drop, the environment changes significantly, or a serious error occurs. Increasing autonomy comes with the risk of greater consequences if the agent is wrong, manipulated, or compromised.
Security leaders must weigh these risks carefully, as 46% of CISOs list AI agents acting outside their intended permissions or being compromised among their top five concerns.
Written by urgent.news from The Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.