Wazuh Bolts Claude and Llama Onto SOC Workflows — Mind the New Attack Surface

Wazuh, the open-source security monitoring and XDR platform, has rolled out AI-assisted reporting and chat features aimed at a familiar SOC problem: too many alerts, too little time to correlate them. The rollout covers three separate integration paths, each with a different trust model — and that difference matters more than the marketing copy suggests.

Three integrations, three trust boundaries

• Wazuh AI Analyst (Wazuh Cloud) — a managed, scheduled service that, per Wazuh's documentation, processes security data through Amazon Bedrock and Anthropic's Claude to generate PDF reports covering key indicators, endpoint histograms, alert volumes, active vulnerabilities, and posture summaries, delivered under Environments > AI Reports. Wazuh states subscription data is not shared with third parties or used to train models.
• Self-hosted Llama 3 via Ollama — for teams with data-residency constraints, Wazuh supports running Meta's Llama locally, with Python scripts vectorizing logs into a FAISS store and a LangChain-powered chatbot querying them entirely on-premises.
• Claude 3.5 Haiku in the dashboard — a chat interface surfaced through the OpenSearch Assistant, using Claude 3.5 Haiku hosted on Amazon Bedrock, described in Wazuh's own blog as guidance for findings and configuration tasks.

Why this matters beyond alert fatigue

The stated goal — cutting through alert volume and manual dashboard-hopping — is real and worth taking seriously; correlation fatigue is a genuine driver of missed detections. But bolting a chat or summarization layer onto a SIEM changes what that LLM is reading. Alert data, DNS queries, HTTP headers, and process command lines are all attacker-influenced fields. A model summarizing or answering questions about that data is, by definition, processing untrusted input alongside its instructions — the same shape of problem as prompt injection in any other LLM application, just relocated into the security stack itself.

That doesn't make these features unsafe to deploy. It does mean the SOC assistant deserves the same scrutiny an organization would apply to any other agent that ingests untrusted content: what happens if a crafted User-Agent string or a malicious hostname is designed to manipulate the model's summary or its recommended remediation? Wazuh's architecture — Bedrock isolation, no permanent processing storage, no training on customer data for the managed service — addresses data-handling risk, not necessarily instruction-following manipulation from within the log stream.

The practical trade-off

The managed-versus-self-hosted split is the more immediate decision for most teams. Wazuh Cloud's AI Analyst offloads inference to Bedrock/Claude with stated no-retention and no-training guarantees, which suits organizations comfortable sending security telemetry to a third-party cloud pipeline. The Llama 3/Ollama path keeps everything on-premises at the cost of running and maintaining the model yourself — the right call for regulated environments where security data cannot leave the perimeter at all, regardless of vendor assurances.

The bottom line

AI-assisted SOC tooling is arriving fast, and Wazuh's multi-path approach — managed, self-hosted, and embedded assistant — is a sensible way to let teams pick their risk tolerance. The open question for any adopter is less "does the report look good" and more "what happens when the input data is adversarial," since in a SOC, some of it always will be.