Loading...
Webinar
Hiring or Firing — The Next Step for AI-Augmented SOC Teams
Webinar
Hiring or Firing — The Next Step for AI-Augmented SOC Teams
Hiring or Firing — The Next Step for AI-Augmented SOC Teams
Hiring or Firing — The Next Step for AI-Augmented SOC Teams
Loading...
Drawing on over two decades of experience delivering security services to global enterprises, Ashish outlines a structured, practical model for building out highly available, customized, and cost-effective AI-enabled security operations.
Security operations teams are eager to have intelligent software agents take on increasing portions of investigative and operational workloads. While much of today's conversation focuses on the use cases for such agents, like alert triage or change assessment, it is equally important to consider how these agents should operate as they scale. A single agent may provide immediate benefits, but adding dozens, hundreds, or even tens of thousands of agents introduces new coordination and security challenges. Safe, compliant, cost-effective adoption requires a structured model for operational maturity.
This essay describes a four-phase operational maturity model to help organizations adopt agentic AI, based on my work leading Cyber Delivery for Infosys. This model is grounded in real implementation patterns: near-term wins, the need to manage permissions across agents, the need to ensure resilience when agents are business-critical, and an expected multi-vendor agent ecosystem.
A good starting point is a narrowly scoped agent supporting the SOC. This agent performs inline investigations of alerts and hands results to human analysts for review, approval, and further action. The agent is designed to "ask for help" when confidence is low or when policy requires human involvement. Phase 1 remains intentionally simple, avoiding cross-functional automation. It focuses on a single workflow, often deployed in one region or business unit at a time.
At this phase organizations should begin collecting the agent's decision traces as the basis of their own security data set. As the agent executes investigations, summarizes incidents, identifies misconfigurations, and interprets context, it generates a wealth of structured reasoning data that can ultimately help the enterprise develop an internal model tailored to its technology stack, processes, and threat landscape.
Leaders can evaluate Phase 1 performance along three dimensions: agent stability, agent accuracy, and agent security, testing for bias and validating the explainability of AI outcomes. These baseline requirements ensure that the agent behaves predictably and safely before its responsibilities expand. The target should be to automate around 80% of SOC alert processing workload before moving to Phase 2.
Phase 2 introduces a broader agent ecosystem. Additional or expanded agents come online to support adjacent functions like vulnerability management, SecOps change workflows, or threat intelligence correlation. Each agent or function continues to be evaluated on stability, accuracy, and security as in Phase 1. Phase 2 continues the building of the organization's security data set with the objective of enabling a purpose-built Security Language Model, as described below.
Phase 2 adds a critical new component to the environment: an identity fabric that governs how agents authenticate, what permissions they receive, and how they may communicate with one another. This identity fabric becomes the central governance of the ecosystem. It provides least-privilege entitlements, lifecycle controls to prevent sprawl, and the ability to isolate any agent that behaves unexpectedly. The fabric ensures that cross-agent cooperation is safe, observable, and reversible.
While much of today's conversation focuses on the use cases … it is equally important to consider how these agents should operate as they scale.
Change Validation is a good example of the complex, multi-stage workflows that are possible in Phase 2, in this case to catch and minimize the impact of an unauthorized change:
This workflow can be executed by a single multi-function agent or cooperatively by a collection of specialized agents working across different teams, applications, and infrastructure.
By the end of Phase 2 organizations should see that 60–70% of the security operations tasks previously performed by humans can be completed by agents, with humans stepping in for exceptions. The progression from Phase 1 to Phase 2 should be possible in 9-12 months.
Once agents become integral to daily operations, they must be engineered for resilience. Traditionally, business continuity planning focuses on infrastructure redundancy with multiple sites, redundant network paths, and failover capabilities. Level 3 extends this thinking to AI agents.
Organization should expect that both humans and agents may at some point be unavailable and prepare for smooth substitution in either direction. If an agent fails, is quarantined, or loses access to its toolchains, humans can take over its workflows. If human operators are unavailable, for example, during night shifts or unexpected workforce disruptions, agents can maintain minimal viable operations without waiting for manual intervention.
To support this capability, enterprises deploy agents across multiple fault domains, such as two cloud availability zones or mirrored data centers. This will often need to consider data-sovereignty and compliance requirements. They may designate secondary "battalions" of agents that remain on standby until a trigger awakens them to preserve continuity. A site under DDoS attack, for instance, might have its primary agents consumed with containment efforts, triggering the activation of a second cohort in another region that can maintain standard SecOps workflows.
Level 3 also strengthens response to major incidents with better summarization and context delivery. Highly available agents can produce actionable digests of recent events such as configuration changes, correlated alerts, and topology shifts that dramatically accelerate human decision-making. An eight-hour downtime could conceivably be reduced to a fraction of that, not through defensive heroism but through reliable, context-rich coordination between agents and humans.
The fourth level represents an environment where agents developed by different vendors can interoperate seamlessly regardless of their underlying reasoning engines. In this model, agents might come from firewall vendors, cloud providers, managed security service partners, or specialized AI companies. They may rely on LLMs or adopt future reasoning technologies not yet invented. Each agent may be optimized for its own domain, but they must operate under a common set of policies and communication standards. Enterprises must therefore design for heterogeneity, ensuring that their identity fabric, authorization model, and operational governance remain vendor-neutral and that vendor diversity does not lead to fragmentation or security gaps.
Ultimately, Level 4 is less about technical complexity and more about strategic resilience. The enterprise positions itself to continuously adopt emerging AI capabilities without destabilizing its security operations and losing is accumulated knowledge base.
For security operations leaders, the reward of adopting this maturity model extends beyond detection and response. It enhances availability, reduces operational drag, and shortens learning curve across staff transitions. As enterprises progress, their accumulated security experience becomes a living repository of organizational defense knowledge, one that strengthens with every investigation, change validation, and agent interaction. In a landscape where threats evolve daily and technologies shift rapidly, such institutional memory may be the most valuable outcome of all.
While "AI" and "LLM" are almost synonymous today, this is unlikely to be the case going forward. Just as utility companies move from coal to nuclear to solar power, companies will have different technologies powering their AI agents. Organizations should be thinking about the evolution in their use of LLMs.
Organizations begin their AI journey with public large language models that come packaged with commercial agent frameworks. This is the fastest and least disruptive way to evaluate agentic capabilities, requiring no up-front model training and delivering strong out-of-the-box reasoning performance. However, cost and data privacy constraints emerge as adoption grows. LLM tokens frequently account for 50% or more of operating expenses, which becomes unsustainable as agents begin handling high-volume telemetry. Masking or redacting sensitive data keeps it secure but also degrades model performance.
As an alternative, organization should start planning for a Security Language Model (SLM), a private model trained on the organization's own security data, telemetry, and incident investigations. Such a model reduces dependence on public LLMs, improves accuracy and explainability, and preserves institutional knowledge even during staff turnover or vendor transitions. As agents and technologies evolve the SLM provides continuity, capturing the enterprise's understanding of its own environment. By capturing everything agents do from the start, enterprises create an asset that endures beyond today's LLMs and can be adapted to tomorrow's agentic solutions.
Read the full ebook → Security for Winners: The Art of Using AI to Secure Your Company and Get Yourself Promoted
