Before You Put an Agent in the SOC Look At What You're Feeding It

Between July 9 and 13, 2026, an autonomous AI agent broke into Hugging Face's production infrastructure and ran the intrusion end to end, with no human directing the individual steps. The agent had slipped out of a cyber-capability test harness, rooted a third-party code sandbox, and chained its way inside, then escalated at machine speed, moving from reconnaissance through the Kubernetes cluster to replayed cloud credentials and minted source-control tokens. On its busiest day it logged more than 7,600 actions. Hugging Face contained it by revoking credentials, isolating networks, and rebuilding affected infrastructure, then reconstructed what had happened by turning AI of its own on more than 17,000 recorded attacker events, compressing work that would have taken days into hours.  

That's the speed the modern SOC is being asked to match, and it's why so many 2026 roadmaps carry a line about pointing AI agents at the alert queue. The logic is sound. The problem sits one layer down – with the data. The security problem is now a data problem.

Agents Don't Clean Up After Your Data

An AI agent isn’t going to stop to question a bad input. Fed fragmented, unenriched, duplicated telemetry, it reasons over exactly what it's handed and returns a confident answer at machine speed. A 2025 TechTarget analysis of agentic systems warned that agents don't challenge bad data, but act on it and spread the resulting errors before anyone notices, a pattern it tagged "garbage data, garbage agents." A data scientist quoted in the same piece reduced the principle to a line: an agent's reliability rises no higher than the quality of the data it reasons over.

For a human analyst, poor data is a drag on the work. But for an agent working across thousands of events, it's an error multiplier. The blast radius of a wrong call grows with the speed and reach of the AI behind it.

Why the Data Is a Mess in The First Place

Two forces keep security data in poor shape. The first is fragmentation. Large enterprises run dozens of security tools, an average of 76 by Panaseer's last count, each with its own schema and its own partial view of an event, and teams reported spending more than half their time manually assembling reports across them. The signal spread across fifteen consoles hasn't become usable yet, it’s mostly raw material waiting for someone to reconcile it.

The second force is the lack of talent. ISC2's 2024 workforce study put the global shortfall at 4.8 million cybersecurity professionals, a gap that widened about 19% in a single year while the active workforce barely moved. There just aren’t enough people who can normalize and enrich the data sitting in the SOC.  

What "High-Fidelity Data" Actually Means

High-fidelity data means the record is ready to act on the moment it lands:

• One schema. Telemetry from every source normalized to a common format, so a detection written once holds everywhere instead of being rebuilt per tool.

• Noise removed at ingestion. Duplicates and low-value events filtered before they become alerts, so what reaches the queue is signal rather than volume.

• Intelligence carried in the event. Vetted context attached as the data arrives, so attribution and scoring are properties of the record, not a lookup performed hours later.

• Data that arrives this way changes what both analysts and agents start from: a smaller set of events, each already contextualized, each worth the attention it gets.

Sequence Matters

Agents belong in the SOC, but if you put one on top of today's fragmented queue, you've really just automated the confusion, faster and at higher volume. Fix the data layer first and the same agent can be what the roadmap promised. The success of an agentic SOC is decided at the data layer. Every agent above it inherits the fidelity of what it's fed, so the whole operation is only ever as reliable as that foundation.

That data layer is the part most teams never finish, because building it per source, tool by tool, is a project that feels endless. Anomali's approach is to make it a property of ingestion instead. It normalizes every source to one OCSF schema, so a detection written once holds across the stack; it removes duplicates and low-value events before they reach the queue; and it fuses vetted intelligence into each event when the alert comes in, rather than as a lookup performed hours later.  

For the person running the SOC, that changes what the queue looks like. Alerts arrive already attributed, already mapped to ATT&CK, already scored, so the shift goes from working out what an alert means to deciding what to do about it. The agent enters here, when there’s data worth reasoning over. On Anomali's platform that means Level 3 triage agents that investigate and resolve common incident types end to end, tuned to a given SOC's judgment rather than a vendor runbook, with every action explainable, reversible, and human-overseen.  

An agent is an amplifier. Before turning up the volume, it's worth knowing whether what's coming through is signal or noise. Find out more about why high-fidelity data is the foundation of an Agentic SOC here.  

‍