A staged framework for granting AI agents autonomy in security operations
Security leaders are moving toward autonomous AI agents for low-risk tasks. A new framework breaks autonomy into stages and levels to manage risk effectively.
Security operations centers are preparing to grant artificial intelligence agents significantly more authority over the next year. According to a recent survey by a16z, 36 percent of CISOs at large organizations currently allow AI to act autonomously on low-risk tasks, a figure expected to rise to 60 percent within twelve months. This shift is driven by the urgent need to counter AI-accelerated attacks that outpace human response capabilities.
What happened
The primary driver for this transition is the scarcity of analyst time. In the same a16z survey, 55 percent of respondents identified AI-accelerated attacks as their top concern, noting that teams struggle to keep up with the volume and speed of threats. Every security alert requires pulling evidence from multiple systems, reconstructing events, and determining relevance, which consumes limited human resources. While demand for automation is high, caution remains prevalent for high-stakes actions. Only 1 percent of leaders are comfortable with full autonomy for most actions today, and just 7 percent expect to be so in a year.
Current discussions often frame autonomy as a binary choice between AI suggesting actions and AI executing them. This oversimplification causes many programs to stall because it fails to account for the varying risk levels of different security tasks. Experts argue that autonomy should not be a single slider but a set of decisions made for each stage of the workflow. Teams that succeed treat autonomy as something an agent earns through demonstrated reliability and evidence, rather than a blanket permission granted at the start.
How it works
To manage this complexity, security leaders can apply a framework originally developed for aviation automation, which divides tasks into four functions: information acquisition, analysis, decision selection, and action implementation. In a security context, these map to evidence collection, analysis and verdict, response recommendation, and response execution. Each function can have a different level of automation. For instance, a team might fully automate evidence gathering and analysis, where errors are cheap to correct, while keeping human oversight for execution, where mistakes can disrupt business operations.
Researchers at the University of Washington further refine this by defining five levels of agent autonomy based on the human role. Level 1 involves the human as an operator who directs every step. Level 2 is collaboration, where human and agent plan together. Level 3 casts the agent as a consultant that leads but checks in for expertise. Level 4 allows the agent to work independently, seeking approval only for risky cases. Level 5 is full autonomy, with the human acting merely as an observer who can hit a stop button. This vocabulary helps teams specify exactly how much independence an agent has at each stage of the security workflow.
Key details
- 36% of CISOs currently allow AI autonomy for low-risk actions, rising to 60% in one year.
- 55% of security leaders worry about AI-accelerated attacks outpacing their response teams.
- Autonomy should be split across four stages: collection, analysis, recommendation, and execution.
- Five autonomy levels range from human operator (L1) to human observer (L5).
- Decisions on autonomy levels should rely on reversibility, blast radius, and evidence strength.
- 46% of CISOs fear AI agents acting outside intended permissions or being compromised.
Why it matters
For software engineers and security leads, this framework provides a practical method to scale AI adoption without introducing unacceptable risk. By automating low-risk, high-volume tasks like log querying and indicator enrichment, teams can free up analyst hours for complex investigations. This approach prevents the bottleneck of requiring human approval for every minor action, which often leads to "rubber stamping" where reviewers approve requests without proper verification due to queue pressure.
However, granting autonomy also increases the potential blast radius if an agent is compromised or makes an error. The survey highlights that 46% of CISOs are concerned about agents acting outside their permissions. Treating the AI agent as a privileged identity with scoped permissions, dedicated logging, and strict monitoring becomes essential. This ensures that as autonomy expands, the ability to detect and reverse harmful actions remains intact, protecting the organization from both external threats and internal AI failures.
What you can do
- Inventory all actions your AI agent can perform and map them to the four workflow stages.
- Score each action using a rubric based on reversibility, blast radius, and evidence quality.
- Automate evidence gathering and analysis aggressively, as these carry the lowest risk.
- Require clear evidence trails for any action requiring human approval to prevent rubber-stamping.
- Define specific graduation and demotion criteria for autonomy levels before starting pilots.
- Govern the AI agent like a privileged user, with unique identities, scoped permissions, and audit logs.



