AI in the SOC: what it automates today and what stays human
Where AI helps in a SOC (summaries, enrichment, triage) and where it fails (context, false positives, decisions). With limits and my own examples.
Written by Eduardo Olivares, SOC analyst in Barcelona.
- ai
- soc
- security copilot
- owasp llm
- nist ai rmf
- soc analyst
People who search for "ai soc analyst" usually want to know one of two things: whether this takes their job, or whether it is useful for anything today. Here I separate what the vendors' documentation says these tools do from what still needs a person, using the limits the vendors themselves acknowledge.
Short answer: AI helps with bounded tasks on data that has already been collected: summarizing alerts, turning a question into KQL, giving context on indicators and drafting reports. It fails where you need to know the business, and its output can be inaccurate or incomplete, so someone has to check it before acting. Microsoft says so in the Security Copilot documentation: its agents are designed to augment human capability, not replace it, and the customer remains responsible for reviewing the results.
What "AI-assisted SOC analyst" means
There are two usage models and it is best not to mix them up. The copilot answers what you ask and you decide what to do with the answer. The agent processes signals from the environment, generates recommendations and, according to Microsoft's definition, can run bounded actions within the configured permissions, with approval from a user or administrator. Microsoft describes them as semi-autonomous flows with human oversight.
The difference changes the risk. A wrong copilot wastes your time; a wrong agent with write permissions changes things in the environment. The OWASP Top 10 for LLM applications (2025 edition) covers this as LLM06, Excessive Agency.
Tasks where it helps
Summarizing alerts and incidents
This is the best-documented use. According to Microsoft Learn, Security Copilot generates summaries of complex alerts and correlates signals across Microsoft Defender XDR, Microsoft Sentinel and other integrated products. A summary is useful because you can check it against the original alert in seconds. My approach is to read it as an index and open the alert before taking anything as settled. For the order in which to look at what fires, I wrote up alert triage in a SOC.
Enriching IOCs and CVEs, and writing queries
The Security Copilot documentation describes plugins that search Microsoft Defender threat intelligence articles and profiles, threat analytics reports and vulnerability publications. It also translates natural language into KQL queries for hunting.
Part of the enrichment does not need a model. Watching CISA's KEV catalog can be solved with a script, and I covered that in this post. The priority of a CVE depends on more than its score, as I explain in prioritizing CVEs without relying only on CVSS. With KQL there is an explicit warning from Microsoft: the generated code can look valid without being correct or reflecting what you asked for, and it should not reach production without validation. Their example is checking that the time range in the query matches the one you asked for in the prompt.
Drafting reports
Microsoft includes among its use cases reports for different audiences, from technical to executive summaries. Collecting the data is the mechanical part and is solved without a model, as in the weekly SOC report with Python.
Where it fails or needs human review
Lack of business context
Microsoft states that responses depend on the data sources, the integrations and the context the user provides, and that if relevant data is missing the result may be incomplete or inaccurate. A model does not know that a server is critical for invoicing if nobody has told it.
Hallucinations and traceability
OWASP lists misinformation as LLM09 (Misinformation). Microsoft admits that Security Copilot can produce inaccurate, incomplete or outdated answers, and defines overreliance as accepting incorrect outputs because the errors are hard to detect. For an analyst, it warns, that can translate into missed threats or wrong conclusions. As a mitigation, the tool cites sources and shows the intermediate steps, and you should demand that from any product.
A small example of my own on why it pays to measure. In the post on disagreement between classifiers I tested a more sophisticated idea against a simple baseline, using 7,822 emails. The text-only model reached an accuracy of 98.17%, almost the same as the full ensemble, and the elaborate idea lost to it. The corpus covers 2004 to 2007, so those figures do not carry over to a current SOC. What does carry over is the method: compare against a simple baseline before adding complexity.
Sensitive data going to third parties
OWASP includes sensitive information disclosure as LLM02 and prompt injection as LLM01. The second has a concrete consequence in a SOC: the email or script you are analyzing was written by the attacker, and it may contain instructions aimed at the model. On data, Microsoft states that in Security Copilot prompts and responses are treated as customer data and are not used to train the foundation models. That guarantee belongs to that product; for any other, you have to ask for it in writing. One alternative is for the data never to leave the machine, as in the phishing classifier that runs in the browser, where the URL does not leave the device.
How to evaluate a tool: checklist
- Data: what leaves your environment, who sees it, how long it is kept and whether it is used for training.
- Permissions: what the agent can do, under which identity and whether least privilege is applied. Microsoft recommends configuring RBAC and agent permissions on that basis.
- Auditing: whether it shows the steps and the sources, and whether there is a record of who launched what.
- Error rate: ask for a test with your own cases, counting false positives and false negatives. A vendor figure on data that is not yours is of no use to you.
- Cost and capacity: in Security Copilot, capacity is measured in security compute units (SCU), and Microsoft recommends monitoring their consumption.
- Risk framework: NIST AI RMF 1.0 (published on January 26, 2023, voluntary) organizes management into four functions: Govern, Map, Measure and Manage. The OWASP Top 10 for LLM 2025 works as a technical checklist.
Which skills still matter for getting into the SOC
To check an AI output you have to know how to do the work by hand. That includes reading logs, understanding the order of triage, writing and reading KQL and knowing how the environment you protect works. If you are starting out, what a SOC analyst does and certifications to get started are the place to begin.
Frequently asked questions
Will AI replace the SOC analyst?
I cannot predict the future. What Microsoft's documentation says today about Security Copilot is that it is designed to augment human capability and that customers remain responsible for reviewing and validating its results.
What is an AI agent in a SOC?
According to Microsoft's definition, software that processes signals from the environment, analyzes data and generates recommendations, and that can run bounded actions within the configured permissions, with approval.
Can I paste incident data into a public chatbot?
It depends on your organization's policy and the contract with the provider. OWASP includes sensitive information disclosure among the risks of LLM applications. Without explicit authorization, do not paste customer or user data.
Which frameworks help evaluate the risks of an AI tool?
NIST AI RMF 1.0, which is voluntary, and the OWASP Top 10 for LLM applications in its 2025 edition. In July 2024 NIST also published a specific profile for generative AI (NIST AI 600-1).
Do I need to know KQL if an AI generates it?
Yes. Microsoft warns that generated code can look correct without being so and that it must be validated before use in production. To spot an error in the query you have to be able to read it.