Security teams are evaluating agentic AI because security operations need faster triage, investigation, and response. The appeal is clear: AI agents can work across alerts and data sources at machine speed. However, speed alone does not guarantee better decisions.For defensive AI, context determines whether an agent understands what it is seeing, why it matters, and what action is appropriate. This article explains why context is central to defensive agentic AI, what can go wrong when context is poorly managed, and why some organizations may prefer AI-supported decision intelligence before moving toward autonomous action.
What Agentic AI Means in Security
A large language model responds to the information placed in its prompt and surrounding conversation. Agentic AI goes further, in line with its design to pursue a goal, use tools, gather information, and iterate toward an outcome.In cybersecurity, that outcome might be alert triage, incident investigation, vulnerability prioritization, or a response recommendation. An agent may review endpoint alerts, identity signals, asset data, threat intelligence, and other sources before deciding what should happen next.Context goes far beyond a prompt, and can include:
Asset inventory and business criticality.
Security controls and configuration data.
Exposure and vulnerability information.
Identity and access activity.
Endpoint, network, and cloud telemetry.
Threat intelligence.
CMDB records.
Alert history and investigation notes.
Policies that define what the agent can and cannot do.
This is what makes context engineering, the curating and maintaining of the information available to a model during inference, so essential.Simply put, any instructions, tools, external data, message history, and other inputs will inevitably shape agent behavior.The practical point for security teams is simple. Defensive AI does not need every available signal at once. It needs the right signals to make a decision that genuinely reduces risk.
Why Context Matters to Security Outcomes
The business value of agentic security depends on the quality of decisions. The promise of faster triage, lower analyst workload, and shorter response times can only be met if the AI can separate real risk from routine noise.For Emanuel Salmona, CEO and co-founder at Nagomi Security, an agent is only as good as the context it operates on: “Give it incomplete data and it will still act. Confidently. Quickly. Incorrectly. Automation without verified context is just a faster way to be wrong at scale.”That is the core risk. An agent may produce a clear recommendation even when it lacks the evidence needed to support it.Consider an alert involving a device that appears compromised. If the agent only sees the technical signal, it may recommend immediate isolation. That may be reasonable from a narrow security perspective. But if the device supports a business-critical process, isolation could create operational disruption. The decision is not only about the alert; it is also about the asset, its role, dependencies, and business impact.In other words, too little context creates blind spots. An agent may see an endpoint event without identity data, a vulnerability without asset criticality, or suspicious access without recent administrative changes, and proceed to infer what it does not reliably know.Meanwhile, larger context windows can reduce precision in retrieval and long-range reasoning. Too much context can lead to slower reasoning, loss of focus, loops, or hallucinations as the agent tries to connect loosely related data.This is why “send everything to the AI” is not a reliable strategy. SOCs already receive large volumes of data from tools such as EDR, NDR, XDR, SIEM, SOAR, IAM, threat intelligence platforms, and software bills of materials. To summarize, agentic AI does not remove the challenge of interpreting and using data. It simply moves it into the AI workflow design. So if the agent has incomplete, stale, or excessive context, it may accelerate poor decisions rather than improve security operations.
An Alternative: AI Without Full Agentic Autonomy
The SOC is a natural use case for agentic AI because analysts already work under pressure from alert volume, time constraints, and complex tool environments. Automating parts of triage can help teams move faster and focus analyst time on higher-value work.But do you really need to have an agentic AI system?One alternative is to use AI for decision intelligence without allowing autonomous action. Lanxit’s Security Decision Intelligence Layer is an example of a system that uses AI, but is not an agentic AI system.According to Obbe Knoop, founder and CEO at Lanxit, a human expert can easily pull up relevant context from VPN gateways, identity systems, Active Directory, and take a closer look at the threat intel, the target’s CMDB, and the business structure CMDB data.The AI agent analyzes the data, notes whether there is sufficient context, and provides a recommendation in plain language. However, it does not take autonomous action: the final decision remains with a single expert or a team.The reasoning behind this approach is that agentic AI still isn’t mature enough. While it can help analysts process a vast amount of information, there is a benefit to ensuring that people take charge of high-impact actions.
What Organizations Should Prioritize
Before expanding agentic AI in defensive security, organizations should focus on the conditions that make AI-supported decisions reliable.1. Define the Agent’s Scope
An AI agent should have a clear purpose. Broad goals create ambiguity and increase the chance of poor decisions.For example, a triage assistant should not automatically become a containment engine without separate controls, testing, and governance. The context required for phishing triage differs from that needed for vulnerability prioritization or cloud misconfiguration analysis.Scope determines context. Without scope, teams cannot reliably decide what information the agent needs or what actions it should be allowed to recommend.
2. Curate Context Instead of Aggregating Data
Security teams should treat context as a managed input, not a data dump. Anthropic recommends finding the smallest set of high-signal information that supports the desired outcome.In practice, this means providing enough information for the agent to understand the decision while avoiding unrelated data that may dilute the signal. It also means connecting technical telemetry with business context, such as asset criticality, ownership, user role, and operational dependency.The best context is relevant, current, and aligned to the task.
3. Require Explainable Recommendations
AI-generated recommendations should include the evidence and reasoning behind them. A severity score or short label is not enough for decisions that may affect containment, downtime, escalation, or risk acceptance.Adam Irwin, managing partner at Heligan Strategic Advisory, states: “No board would accept a set of numbers without an audit trail, yet many accept intelligence that shapes approvals and decisions with no method of visibility.”Security teams should expect AI systems to explain:
What evidence was reviewed.
Which sources were used.
What uncertainty remains.
Why the recommendation is appropriate.
What options are available next.
Explainability supports analyst review, accountability, and operational confidence.
4. Keep Human Oversight for High-Impact Actions
Autonomous action may be appropriate in narrow, well-controlled scenarios. But high-impact actions, such as device isolation, account suspension, or service shutdown, require stronger governance.Ultimately, it’s up to organizations to decide where AI can recommend, where it can automate with approval, and where it should not act without human review.
Strategic Takeaway
Defensive agentic AI can help security teams operate faster, but its value depends on the quality of its decisions, which, in turn, depend on context.Too little context leaves the agent without the information needed to judge risk accurately. Add too much of it, and it will create noise, reduce focus, and weaken reasoning. The right balance is deliberate context management: enough high-signal information to support the task, without overwhelming the agent or obscuring the decision.For many organizations, the practical path is gradual. Use AI to gather context, explain findings, and support analyst decisions before allowing autonomous action in sensitive workflows. Agentic AI may become more capable, but business and operational trust will depend on disciplined context design, transparent reasoning, and clear boundaries for human oversight.
