The AI-augmented SOC: promises, limits and the choice of a local sovereign AI

Up to date as of October 31, 2026. This article complements our analyses of the AI Act and of AI agent security.

Rising alert volumes, overstretched teams, scarce skills: the SOC is one of the areas where the most is expected from AI. The promises are many. The reality is more nuanced: AI can save analysts precious time, provided it is well chosen, well framed and well hosted.

What AI really brings

Summaries of alerts and incidents: a readable summary, a timeline, the key elements highlighted, where the analyst used to go through dozens of events.
Help with qualification: enrichment of indicators, matching with threat intelligence, mapping to MITRE ATT&CK tactics and techniques.
Investigation in natural language: turning an analyst’s question into a SIEM query, then explaining the result.
Writing: incident reports, notes for the crisis team, first drafts of detection rules, always reviewed by an expert.
Knowledge transfer: an assistant that explains an event or an attack technique helps junior analysts build their skills.

In 2025, ANSSI also published a market study on AI for incident detection and response. It distinguishes predictive AI, mostly used for detection, from generative AI, which is gaining ground in incident qualification, and highlights the sensitive question of where language models are hosted.

The limits to be aware of

Hallucinations: a model can produce a plausible but wrong explanation. Every conclusion must point to the facts it is based on.
Prompt injection through logs: an attacker can slip instructions into data they control and that the model will read, such as an email subject, a file name or an HTTP header. The logs being analyzed are untrusted data (see our analysis Securing AI agents in the enterprise).
Confidentiality: security logs contain personal data, information about the architecture and sometimes secrets. Sending them to an external service is not a trivial matter, with regard to both the GDPR and trade secrets.
Dependence: a SOC that hands everything over to AI loses skills, and becomes dependent on a supplier, its prices and its product changes.
No measurement: without indicators defined in advance, there is no way to know whether AI really saves time or simply shifts the work.

Our choice: a sovereign AI that runs locally

At Abakus Sécurité, we chose a sovereign AI that runs locally. The models run in a controlled environment and query the SOC tools directly, without security data ever leaving that environment.

In practice, our approach relies on three building blocks:

open models run locally: models with 35 billion parameters (“35B” size), chosen and evaluated according to the use cases;
“skills”: specialized capabilities that tell the assistant how to use a given SOC tool (query the SIEM, look up an incident, enrich an indicator), with rights limited to what is necessary;
a local environment that keeps control of the data, of the model versions and of the traceability of the exchanges.

This choice directly addresses the limits described above: the logs stay within the SOC environment, the models’ behavior does not change with a supplier’s updates, and every interaction can be traced and audited.

We apply it with a few simple principles:

read-only by default: the assistant looks things up and makes suggestions, the analyst decides and acts;
sourced answers: every conclusion points to the events, queries or incidents that support it;
analyzed content treated as untrusted, to limit the risk of injection;
evaluation on real cases before each new use case, new model or new skill.

Bringing AI into the SOC, step by step

Choose two or three specific use cases, for example incident summaries or help with writing queries.
Define the indicators before you start: qualification time, rate of false positives handled, analyst feedback.
Decide on hosting according to the sensitivity of the data: local execution, trusted cloud (ANSSI’s SecNumCloud qualification) or public service, with a risk analysis for each option.
Start in read-only mode, with systematic validation by an analyst.
Log the prompts, the answers and the decisions, and make them auditable.
Test robustness: hallucinations, prompt injection, edge cases, on real anonymized datasets.
Expand gradually and train analysts to work with AI, without losing their expertise.

AI does not replace the analyst: it saves them time on what is repetitive, so they can spend more of it on what requires judgment.

Sources

← All insights