What AI is actually absorbing
The work moving to models first is the work that burned analysts out anyway: tier-one triage of alert floods, log correlation, enrichment lookups, and report drafting. The pattern emerging in SOCs is compression rather than elimination — the same incidents flow through fewer hands, with humans entering later in the chain, at the judgment and authority steps. The roles built entirely on the absorbed tier are the genuinely exposed ones; the judgment tiers above it inherit more influence, not less.
What it cannot take
Accountability does not delegate to a model. Someone decides that a containment action is worth its blast radius, faces the regulator after the breach, designs the threat model for a system nobody has attacked yet, and accepts risk on the organisation's behalf. Adversarial work also resists automation structurally: attackers respond to defences, including AI ones, so the defender's job increasingly includes anticipating how the automation itself will be gamed — work that exists because of AI, not despite it.
The work AI is creating
Every agent an organisation deploys arrives with security work attached: identity and least-privilege design, prompt-injection defence, runtime gates, audit trails, red-teaming, incident response for machine-speed actors. Securing AI is becoming its own discipline while AI-assisted defence reshapes the old one, and practitioners who add it acquire the rarest profile in the field. For a security career today, the practical move is not defensive crouch but expansion: learn how agents work, how they fail, and how they are contained — the demand signal for that combination is visible in any direction you look.