Security News

AI Agents Are Reshaping the SOC, but Entry-Level Analysts Are Being Redeployed, Not Replaced

Infosecurity Magazine · 19 Nov 2026
Key Takeaway If you use or buy AI-driven security monitoring, keep a human approving agent actions at first and widen the agent's autonomy only as its results prove reliable.

Most security leaders are rebuilding their security operations centres (SOCs) around AI agents, but the people at the front line are not disappearing. A Strike48 report, The State of Agentic Security H2 2026, found that 87% of leaders are redesigning their SOCs because of agents, with roles being adapted rather than eliminated.

The security startup polled 534 security and IT leaders at organisations with 200 or more employees and a security operations function in the UK, Canada and the US. It found 74% have now deployed agents in their SOC, up from 36% in May, and 49% are using them in multiple use cases.

First-line (L1 or Tier-1) analysts are mostly being redeployed. Some 49% of respondents are hiring for new skills such as AI detection engineering and automation, and 42% are retraining L1 analysts to supervise agents. A further 38% are moving L1s into investigation roles, while 30% are growing their teams with more entry-level analysts. This eases a common worry that AI would remove a vital career on-ramp, although a separate Swimlane report from September found 24% of SOC analysts feel the technology has limited their ability to build skills.

Trust remains a brake on adoption. Only 30% of teams let agents act with high or full autonomy, while 47% keep humans executing or preapproving actions and 23% allow agents to act within defined boundaries. Strike48 co-founder Tim Leehealy argued the models are mature enough, and that teams reach autonomy by breaking work into narrow jobs and fixed, deterministic workflow steps rather than relying on one long prompt.

Why it matters

Security leaders who assume AI agents will simply replace junior analysts are working from the wrong model: the survey shows the work shifting to supervising agents, engineering detections and investigating. The decision this forces is how much autonomy to grant, since distrust is highest where teams run a single agent on a single use case. Organisations that skip training and governance risk both unreliable automated triage and a thinner pipeline of experienced analysts.

What to do

  • Define which actions an agent may take alone, which need human pre-approval, and write those boundaries down.
  • Break agent triage work into narrow jobs with fixed, deterministic workflow steps rather than one long prompt.
  • Retrain entry-level analysts to supervise agents and move toward investigation work.
  • Review agent outputs against real alert queues before expanding to more use cases.
Running AI agents? Eight controls for securing them, mapped to the ISM ->

Summarised by CISO AI from Infosecurity Magazine, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.