Building an LLM-powered SOC cybersecurity assistant with deception telemetry
In November 2025, Anthropic disclosed a cyber espionage campaign, designated GTG-1002, in which an AI agent carried out most of the tactical intrusion work. Attackers now move at machine speed, and SOC analysts still read alerts one at a time. An LLM cybersecurity assistant for the SOC can close part of that gap, but only if it is built to support the analyst rather than replace their judgment. Building one means choosing bounded tasks, connecting only approved sources, enforcing the analyst’s own permissions at retrieval, citing every statement, and leaving every decision with a person. High-fidelity deception telemetry gives that assistant some of its most useful context.
At a glance
- Definition: An LLM SOC assistant retrieves authorized evidence, summarizes it with citations, and drafts options; the analyst decides.
- Method: Seven steps, from bounded use cases to a monitored limited rollout, each with an owner and a stop condition.
- Decision artifact: A four-row design and evaluation matrix covering summaries, triage, retrieval, and failure handling.
- Deception’s role: Decoy interactions are low-volume, high-fidelity context. In the U.S. Navy ANTX FY25 exercise, Acalvio deception produced 100% true positive alerts and denied 80% of attacker objectives against automated, credential-driven techniques in the exercise.
- Limit: An assistant can be wrong, and retrieved text can carry prompt injection, so grounding and evaluation are mandatory.
What an LLM-powered SOC assistant should do
An LLM-powered SOC assistant helps an analyst understand an investigation faster by retrieving authorized evidence, summarizing it with citations, and drafting next steps for the analyst to approve. It is an investigation aid, not an autonomous security control.
Four elements define it. The analyst task is specific: summarize a case, explain an alert, find the relevant playbook, or draft a handoff note. The inputs are authorized sources only, such as SIEM events, case records, asset and identity context, approved playbooks, and threat intelligence. Retrieval is grounded and permission-aware, so the assistant answers from retrieved records rather than from model memory and never shows an analyst data they could not open directly. Every statement carries a citation to the record it came from.
The decision boundary sits with the human. The assistant can recommend containing a host; an analyst decides whether to do it, and any action runs through the existing response tooling and approval process. Teams that blur this line turn a summarization error into an operational one.
Build vs. buy: in-house assistants and vendor copilots
Teams generally choose between building an assistant on their own stack and adopting a copilot embedded in a security platform. Both are valid, and many SOCs use one of each.
| Factor | Build in-house | Vendor-embedded copilot |
| Data access | Any approved source, wired by the team | Strongest within the vendor’s own data; other sources depend on connectors |
| Integration | Full control, full engineering cost | Ready inside the vendor’s console |
| Provenance and citations | Designed to the team’s standard | Depends on the product; verify before relying on it |
| Retention | Set by the team’s own policy | Set by contract and vendor configuration; review carefully |
| Evaluation | Team builds its own test sets | Vendor evaluation plus the team’s own acceptance testing |
| Oversight | Team defines approval gates | Product defines gates; check they match SOC policy |
| Maintenance | Models, prompts, and retrieval need ongoing care | Vendor maintains; the team tracks changes |
Whichever route a team takes, it should verify a named product’s capabilities against its own requirements rather than assuming two copilots behave the same way.
Design a grounded and safe SOC assistant
A grounded SOC assistant answers only from evidence it can cite, and a safe one cannot act or reach data beyond what its user is allowed to see. The seven steps below are an editorial framework that organizes current practice; they are not a published standard.
| Step | Input | Owner | Output | Review or stop condition |
| 1. Choose bounded use cases | Analyst pain points, case history | SOC lead | Two or three named tasks | Drop any task that requires autonomous action |
| 2. Approve telemetry and sources | Data inventory, privacy review | SOC lead and data owners | Approved source list | Exclude sources without an owner |
| 3. Enforce retrieval permissions | Identity and access model | Security architect | Permission-aware retrieval | Block launch if any test returns unauthorized data |
| 4. Expose evidence citations | Retrieval design | Engineering | Answers linked to source records | Reject answers without citations |
| 5. Evaluate incorrect and adversarial outputs | Test cases, injected content | Evaluation owner | Error and abstention results | Stop if unsafe inferences pass review |
| 6. Require analyst decisions | Response workflow | SOC lead | Approval gate on every action | Never bypass for speed |
| 7. Monitor a limited rollout | Pilot group, feedback log | SOC lead | Correction and usage review | Pause on repeated grounding failures |
Acalvio’s data science team described its own early build in 2024: it collected structured MITRE knowledge in STIX format, converted it to Markdown, stored it in a vector database for retrieval-augmented generation (RAG), and compared several open models, including Falcon-7B, Llama 2 7B, and Mistral 7B Instruct, on answer accuracy, response time, and sensitivity to prompt wording. That pattern of retrieval over a curated corpus, followed by repeated evaluation, still holds.
Using deception telemetry as investigation context
Deception telemetry is useful LLM context because each event is low in volume, high in fidelity, and easy to explain: something interacted with an asset that no legitimate user or process should need.
Most SOC telemetry describes authorized activity. A valid login from a plausible host looks the same whether an administrator or an intruder is behind it, so intent is hard to judge in isolation, and an LLM summarizing that activity inherits the same ambiguity. A deception event asks a sharper question: why did anything touch an asset no legitimate process should need? An assistant can retrieve the event, the documented purpose of the deceptive asset, and the related identity, host, and session records, then present them with citations for the analyst to judge.
Hand-built honeypots cannot keep pace with automated, credential-driven intrusions, so automation is the entry requirement. The ShadowPlex Preemptive Cybersecurity Platform deploys and adapts deceptive assets autonomously, and 360 Deception, Acalvio’s three-vector framework, aims to detect, divert, and degrade the attacker. The U.S. Navy ANTX FY25 exercise produced 100% true positive alerts and 80% denial of attacker objectives, against automated, credential-driven techniques in the exercise.
The architecture described here is illustrative. How deception events reach an assistant depends on the integration path a team uses, typically through the SIEM, and should be confirmed with Acalvio before design work begins. Deception runs alongside SIEM, EDR, IAM, PAM, and XDR rather than replacing them, and authorized scanners and red teams can reach deceptive assets too, so the analyst still corroborates every event. The same principle can protect AI systems themselves; see how Acalvio applies deception to agentic AI runtime protection.
SOC assistant design and evaluation matrix
A useful assistant design names, for each task, what the assistant reviews, what its output can and cannot establish, and who acts on it. The rows below are illustrative planning prompts, not a verified deployment or customer result.
| Area | What to review | Evidence and limitation | Owner and next step |
| Case summary | Authorized incident records | Every statement traces to a source | Analyst corrects and approves |
| Event triage | Validated event and asset context | An observation is not intent | Analyst sets priority |
| Knowledge retrieval | Approved playbooks | Check access, freshness, and relevance | Content owner validates |
| Failure handling | Missing evidence or injected text | Test abstention and unsafe inference | A person retains control |
Each row establishes what the assistant may contribute. None establishes that an answer is correct; that takes the analyst’s review and, for anything new, the content owner’s sign-off.
Hypothetical assistant-assisted investigation
This example is hypothetical. It shows the reasoning, not a real customer event.
Objective: Triage an alert that a deceptive service account credential was used to request access to a file server.
Cited retrieval: The analyst asks for a summary. The assistant returns the deception event, the account’s documented purpose as a deceptive asset, and the source host’s ownership record, each with a citation.
Mistaken inference: The assistant adds that “lateral movement is confirmed across three hosts.” Only one host appears in the cited records. The analyst rejects the line and flags it for the evaluation set.
Missing data: EDR telemetry from the source host is not in the approved sources. The correct response is “no endpoint evidence available,” and the assistant is tuned to say so rather than speculate.
Prompt injection: A ticket comment attached to the host reads like an instruction to mark the alert benign. The assistant treats it as data, quotes it, and flags it. This is the indirect form of LLM Prompt Injection that MITRE ATLAS catalogs as AML.T0051.001.
Decision: The analyst corroborates with identity logs, confirms no authorized test was scheduled, and escalates through the incident response process. The assistant drafts the handoff note; the analyst edits and sends it.
Implementation and operational guardrails
An LLM assistant in the SOC touches sensitive data and influences response decisions, so it needs the same controls as any privileged tool.
Before rollout, confirm this readiness checklist:
- Least privilege: the assistant’s service identity can read only approved sources and cannot take actions.
- Source permissions: retrieval enforces each analyst’s own access rights.
- Sensitive-data protection: personal data, credentials, and secrets are masked or excluded from prompts and logs.
- Audit logs: every prompt, retrieval, and answer is logged for review.
- Prompt-injection evaluation: test sets include instructions hidden in tickets, emails, and log fields.
- Retention review: prompt and response retention matches policy and any vendor contract.
- Human authorization: every containment or configuration change needs an analyst’s approval in the existing tooling.
Assign each item an owner, an input, a validation step, an escalation path, and a review date. Product-specific integration steps need Acalvio validation.
Evaluate grounding, usefulness, and safety
Evaluate an assistant on whether its answers are correct, grounded, and safe for the analysts who use them. Set a baseline, a calculation method, and a data owner before reporting any figure.
| Measure | What it shows | What it cannot show |
| Task correctness on a labeled test set | Accuracy on known cases | Accuracy on novel incidents |
| Unsupported statements per answer | Grounding quality | Whether cited sources are themselves right |
| Access-control test failures | Permission enforcement | Every possible bypass |
| Appropriate abstentions | Willingness to say “unknown” | Overall usefulness |
| Analyst corrections | Where the assistant errs in practice | Errors analysts did not notice |
Report evaluation limits alongside results. Time saved and accuracy claims need measured data from the team’s own pilot, not vendor or industry figures.
The path forward
A useful LLM cybersecurity assistant in the SOC is narrow, grounded, and accountable: bounded tasks, approved and permissioned sources, cited answers, tested failure modes, and an analyst who makes every decision. Deception telemetry gives it context that is unusually easy to explain, and the analyst still corroborates it. The goal is to stop asking analysts to infer intent from weak evidence when the environment can produce a stronger signal.
See how 360 Deception produces that signal, or read how deception supports hypothesis-led threat hunting.
Frequently asked questions about LLM SOC assistants
Start with two or three bounded analyst tasks, connect only approved sources, and use permission-aware retrieval so every answer cites the records it came from. Evaluate the assistant against incorrect, missing, and adversarial inputs before a limited rollout, and keep every action under analyst approval.
The best fit depends on which data a team needs the assistant to reach, its retention and provenance requirements, and whether it wants to maintain its own models and retrieval. Compare in-house builds and vendor copilots on those criteria, then run acceptance tests on the team’s own cases.
An assistant can retrieve a deception event, the deceptive asset’s documented purpose, and related identity and host records, then summarize them with citations. Because no legitimate process needs a deceptive asset, the event is in a high-fidelity context, though the analyst still corroborates it.
Ground answers in retrieved records, require a citation for every statement, and reward the assistant for saying “no evidence available.” Track unsupported statements and analyst corrections, and feed every error back into the test set.
Severity, escalation, containment, account changes, and any action that touches production systems. The assistant can draft and recommend; an authorized person approves and executes through existing tooling.
