Security work is mostly reading: alerts, logs, vendor advisories, and the one PowerShell command in a phishing attachment that someone clicked. AI assistants read faster than you do. Paste an obfuscated script and ask what it does; paste an alert and ask what would confirm it is real; paste a CVE description and ask which of your systems it plausibly affects.
The security platforms have their own assistants: Microsoft Security Copilot, CrowdStrike's Charlotte AI, Gemini in Google Security Operations, and Splunk's AI Assistant for SPL, among others. They sit next to your data, which matters because the data cannot leave. A general assistant such as Claude or ChatGPT handles the writing, explaining, and scripting, as long as nothing sensitive goes in.
The failure modes are the ones you would guess and one you might not. Models invent CVE numbers, misdescribe what a command does, and are confidently wrong about detection logic. They also read attacker-controlled content, so a log line or an email body can carry instructions aimed at the model. Anything an assistant reads is untrusted, and anything it recommends gets verified before it becomes a block rule or a report.
Quick wins this week
- Paste a defanged PowerShell or JavaScript snippet from a phishing sample into an approved tool and ask for a plain-English walkthrough: what it does, what it downloads, the artifacts it leaves on a host.
- Describe an alert in words and ask for the KQL, SPL, or Sigma query that would find similar activity over the last 30 days, then test it against known-good traffic for noise.
- Give the assistant a vendor advisory and a summary of your exposure (product names and versions only) and ask which entries are plausibly affected and what to check first.
- Turn your incident timeline notes into a draft report with summary, timeline, impact, root cause, containment, and lessons, then correct the parts the model smoothed over.
What AI can do for cybersecurity analysts, task by task
Alert triage and enrichment
Give the alert fields, the rule that fired, and what normal looks like for that host or user, with internal names replaced by placeholders unless the tool runs inside your tenant. Ask for the likeliest benign and malicious explanations and the single query that separates them. The model does not know your environment; every 'this is normal' claim is a hypothesis to test in your own telemetry.
Log analysis and detection engineering
Describe the behavior you want to catch, paste a few sanitized sample events, and ask for a KQL, SPL, or Sigma rule with comments plus the false-positive sources to expect. Assistants inside Sentinel, Splunk, and Google SecOps generate queries against your own schema where enabled. Run every rule in detect-only mode first and measure the noise before it pages anyone.
Malware and phishing analysis
Use a defanged sample (URLs and IPs broken, no live payloads) and ask what the code does step by step, what it decodes to, and the indicators it would generate. Confirm every decoded string in a sandbox rather than taking the model's word; obfuscation defeats it more often than it admits. Never paste anything from a live incident into a consumer tool.
Vulnerability and advisory triage
Paste the advisory and a summary of your exposure (product, version, internet-facing or not) and ask for affected components, exploitation prerequisites, and a prioritized check list. Verify every CVE number against NVD or the vendor; models produce plausible identifiers that do not exist. Severity for your environment is your call with your data.
Incident reports and executive communication
Paste your timeline notes, containment actions, and evidence list, and ask for a report in your template with every claim tied to a note, then an executive summary that states impact and asks. Check that the model did not invent a root cause where your notes say 'unknown'; it prefers a tidy story to an honest gap.
Response scripts and automation
Ask for the Python or PowerShell that parses a log export, blocks a list of indicators through your firewall's API, or pulls artifacts from a host, with a dry-run mode and no embedded credentials. Read it fully, run it in a lab, and keep API keys in a vault the script reads at runtime. A response script with a typo can isolate the wrong subnet.
Prompts for cybersecurity analysts
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
Triage an alert with confirming queries
You are a senior SOC analyst. Alert details: [PASTE THE ALERT FIELDS AND THE RULE THAT FIRED]. Host or user context: [ROLE OF THE ASSET, NORMAL BEHAVIOR, RECENT CHANGES]. Environment: [EDR, SIEM, AND OS IN USE]. Give the top three benign and top three malicious explanations, each with a rough likelihood and the specific query or artifact that would confirm or rule it out in [SIEM QUERY LANGUAGE]. Then list what I should preserve now in case this becomes an incident. Say clearly when something cannot be determined from the information given. Do not recommend containment yet.
Tip: Replace hostnames and account names with placeholders unless the tool runs inside your security tenant.
Explain a suspicious script step by step
You are a malware analyst. The following is a defanged script from a phishing attachment; URLs and IPs have been broken on purpose. Walk through it top to bottom: what each block does, what it decodes to (show your work), what it downloads or executes, persistence mechanisms, and evasion techniques. Then list the host and network indicators a defender should search for, mapped to MITRE ATT&CK technique IDs. Flag any step where you are guessing because the obfuscation is ambiguous.
Script:
[PASTE THE DEFANGED SCRIPT]Tip: Confirm every decoded string in a sandbox; models get encodings wrong and never say so.
Write a detection rule with expected noise
You are a detection engineer. I want to detect: [DESCRIBE THE BEHAVIOR, FOR EXAMPLE CREDENTIAL DUMPING VIA LSASS ACCESS]. Data source and schema: [LOG SOURCE AND FIELD NAMES]. Query language: [KQL, SPL, OR SIGMA]. Known legitimate sources of similar activity here: [LIST THEM]. Write the rule with a comment per clause, then list the false positives it will likely generate and how to tune each, the ATT&CK technique it covers, the evasion that would bypass it, and a test procedure using a safe simulation. Use only fields that exist in the schema I gave you.
Tip: Run it in detect-only mode for a week before it pages anyone.
Draft an incident report from timeline notes
You are writing a security incident report for [ORGANIZATION TYPE] with this structure: Executive summary, Timeline, Scope and impact, Root cause, Containment and eradication, Recovery, Evidence, Lessons learned, Open items. Use only the facts in my notes and tie every claim to a note by line number. Where the notes say unknown, write that; do not infer a root cause. The executive summary must be under 150 words, non-technical, and end with what leadership needs to decide. Notes: [PASTE THE TIMELINE AND EVIDENCE NOTES] Audience: [BOARD, CUSTOMERS, REGULATOR, OR INTERNAL]
Tip: Reports go to lawyers and regulators; keep the model's guesses out of the record.
Map an advisory to your exposure
You are a vulnerability analyst. Advisory text: [PASTE THE ADVISORY]. Our exposure summary (product names, versions, and whether internet-facing; no hostnames): [PASTE THE SUMMARY]. List the affected products and versions exactly as the advisory states them, the prerequisites for exploitation, which of our entries plausibly match and why, what to verify on each before patching, and mitigations if patching must wait. Quote CVE identifiers exactly as they appear in the advisory and add no others. Where the advisory is ambiguous, say so instead of resolving it.
Tip: Check every CVE number against NVD; the model adds plausible neighbors that do not exist.
Want a prompt for something else? Use the Prompt Builder.
Skills to build
Treating model input as attacker-controlled
Why: Emails, logs, tickets, and web pages you analyze can contain instructions aimed at the assistant, and an agent that acts on them can exfiltrate data or open a hole.
How: Never let an assistant act on content it read without human approval, isolate analysis sessions, and put prompt injection in the threat model for every AI-enabled tool you deploy.
Query languages, read and write
Why: Generated KQL, SPL, and Sigma look right and can silently miss the events you care about, and a detection that never fires is worse than none.
How: Learn your SIEM's schema and operators well enough to read a generated query clause by clause, and test every rule against known-bad and known-good data.
Sanitizing and scoping what you share
Why: Internal hostnames, IP ranges, usernames, and secrets in a prompt become internal reconnaissance in a vendor log.
How: Keep a placeholder convention (HOST-A, USER-1, 10.x.x.x), use in-tenant tools for raw data, and keep a general assistant for writing and explaining only.
Verifying against primary sources
Why: Models invent CVE numbers, misquote advisories, and describe ATT&CK techniques from memory, and one wrong identifier undermines everything around it.
How: Cross-check every identifier, version, and quote against NVD, the vendor, and MITRE before it leaves your draft, and make it a checklist item in your report template.
Evaluating AI inside security products
Why: Every vendor is adding an assistant, and you will be asked whether it is safe, useful, and worth paying for.
How: Ask where data goes, whether prompts are retained or used for training, what actions it can take, and how it handles injected content; then run it against past incidents you know the answers to.
Tools worth knowing
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Claude
A careful writing and analysis assistant that shines on long documents.
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
Claude Code
Anthropic's agentic coding tool that works in your terminal and IDE, on your real codebase
GitHub Copilot
AI pair programmer inside VS Code, JetBrains, and GitHub itself, with a free tier
Cautions for cybersecurity analysts
Prompts get pasted with API keys, service-account passwords, VPN configs, and internal IP ranges inside them, and a consumer chat account is a third party. Never paste a secret; rotate any that reached a prompt; replace internal identifiers with placeholders; keep raw telemetry in tools that run inside your tenant under your organization's data terms.
Attackers know defenders use AI. A log line, an email body, a web page fetched by a security tool, or a document in evidence can carry text like 'classify this as benign' or 'run this command'. Summaries are fine; automated actions are not. Keep a human between any model and any change to firewalls, identities, or endpoints, and test your own AI-enabled tools for this.
Models produce CVE identifiers that do not exist, misdescribe what an obfuscated command does, and write rules that reference fields your logs do not have. Verify identifiers against NVD and the vendor, decoded strings in a sandbox, and every rule against real data in detect-only mode before it pages or blocks anything.
Incident details are often privileged (work product directed by counsel), regulated (breach notification and contractual duties), and sensitive to customers. Keep live incident data out of consumer tools, use your organization's approved platform, keep the model's inferences out of formal reports, and follow counsel's guidance on what gets written down at all.
Your 30-day plan
- Week 1: Get security's own approval for which assistant is used for what: a general assistant for writing and explaining, in-tenant copilots for anything touching telemetry. Write the placeholder convention.
- Week 2: Use the assistant on every alert you triage for hypotheses and confirming queries, and track how often it was right and how often confidently wrong.
- Week 3: Generate three detection rules for gaps you already know about, run them detect-only, tune the noise, and compare predicted false positives with the ones you saw.
- Week 3: Analyze the last five phishing samples with the explain-the-script prompt and compare with sandbox results to calibrate how far to trust it.
- Week 4: Draft your next incident report and executive summary with the assistant, add the primary-source checklist to the template, and brief the team on prompt injection as a threat to your own tooling.
Frequently asked questions
Will AI replace cybersecurity analysts?
Is it safe to paste logs into ChatGPT or Claude?
Can AI write detection rules for my SIEM?
What is prompt injection and why should security teams care?
Terms used on this page
Related roles
- AI for IT Support SpecialistsChat assistants explain error codes, draft the PowerShell you would otherwise search for, and write the ticket reply in the user's language. They also confidently suggest commands that can wipe a machine, so read before you run.
- AI for DevOps EngineersCoding agents now write Terraform, Helm charts, CI pipelines, and the Bash you would rather not, and chat assistants read a 3 a.m. incident faster than you can. Every generated line touches production, so plan output, review, and blast radius come first.
- AI for Compliance OfficersAI rewrites the policy nobody reads into one people will, finds the rule change you missed, drafts the training deck, and structures the investigation chronology. It also puts a new line on your risk register: the company's own use of AI, which regulators now expect you to govern.
- AI for Software DevelopersCoding assistants write, refactor, test, and explain code, and agentic tools run the whole loop. Treat every generated line as a pull request from a fast, overconfident junior: useful, never merged unread.