The AI Learning Hub Journal

Prompting for Security Work

Three daily wins, one hard linerole, audience, format, then the raw material — and verify before anything travelsSUMMARISING ALERTS + INCIDENTSTHE PATTERNstate the role and audience,constrain the format — timeline,assets, actions, open questions —then paste the source materialTHE CHECKevery timestamp, hostname, andindicator checked against source— gaps get plausible fabricationDRAFTING COMMS UNDER PRESSURETHE PATTERNgive the facts, the audience,and the desired reader action —let it handle register: execsummary, notice, status updateTHE CHECKfactual claims, liability wording,and the decision to send stayhuman — drafts are discoverableQUERYING DOCS AND STANDARDSTHE PATTERNsupply the document, then askfor gaps, contradictions, orapplication to a scenario —require quoted passagesTHE CHECKun-supplied documents getanswered from stale trainingdata — treat it as a rumourTHE CONFIDENTIALITY LINE — SECURITY PROMPTS ARE SENSITIVE BY DEFAULTknow your tier — consumertools may train on input;enterprise must forbid itsanitise by default —placeholders and maskedIPs keep the valuenever paste credentials,live evidence, or unpatchedvulnerability detailprompt history is a record— assume everything typedpersists somewhereyour organisation's AI use policy should name approved tools per data class — know it before you typeTHE OUTPUT IS A DRAFT THAT SAVES THIRTY MINUTES — NEVER A RECORD YOU CAN SIGNa summariser smooths over gaps with plausible fabrication exactly where the record is thinnest
Summarise, draft, and interrogate documents with structured prompts — then verify every fact, and keep credentials and live evidence out of anything unapproved.

Summarising Incidents and Alerts

The highest-value everyday use of an AI assistant in security work is turning raw material — alert payloads, log excerpts, ticket histories — into structured summaries. The pattern that works: state the role and audience ("you are assisting a SOC analyst; the audience is the incident manager"), constrain the format ("timeline, affected assets, actions taken, open questions — nothing else"), and paste the source material explicitly rather than describing it. Then verify: every timestamp, hostname, and indicator in the output gets checked against the source before it travels anywhere, because a summariser will smooth over gaps with plausible-sounding fabrications precisely where the record is thinnest. The output is a draft that saves you thirty minutes, not a record you can sign.

  • Structure the prompt: role, audience, fixed output format, then the raw material
  • Constrained formats reduce fabrication — free-form narrative invites it
  • Verify every concrete fact against source before the summary leaves your desk
  • Treat the output as a draft that saves time, never as a record of truth

Drafting Communications Under Pressure

Incident communications are high-stakes writing done at the worst possible time, which makes them ideal for AI drafting — and dangerous for AI sending. The assistant is excellent at register: turning a technical timeline into an executive summary, a customer notice, or a status update for legal, each with an appropriate tone you do not have spare cognition to craft mid-incident. Give it the facts, the audience, and what you want the reader to do; let it produce the structure and phrasing. What stays with you: every factual claim, anything bearing on liability or disclosure obligations, and the decision to send. During a live incident with potential legal exposure, drafts can become discoverable records — involve counsel on what gets written where, including in prompt history.

  • Use the assistant for register and structure — executive summary, customer notice, status update
  • Provide the facts, the audience, and the desired reader action; review every claim before sending
  • Liability-relevant wording and disclosure decisions are human work, always
  • Prompt history can be discoverable — during legal-exposure incidents, ask counsel first

Querying Docs, Policies, and Standards

The third daily win is interrogating documents instead of rereading them: paste a vendor security whitepaper and ask what it does not address; paste your own IR policy alongside an incident timeline and ask which steps were skipped; paste a framework excerpt and ask how it applies to a scenario. Asking for gaps and contradictions plays to the model's strength as a tireless close reader, and unlike open-ended questions, answers grounded in a document you supplied are checkable — demand quoted passages with every claim and follow the quotes back. The failure mode is asking about documents you did not supply: the model will describe standards and regulations from training data that may be years stale, with total confidence. If the text is not in the prompt, treat the answer as a rumour.

  • Strongest pattern: supply the document, then ask for gaps, contradictions, or application to a scenario
  • Require quoted passages with every claim, and follow them back to the source
  • Un-supplied documents get answered from stale training data — confidently
  • Version-sensitive material like regulations and standards must always be pasted, never assumed

The Confidentiality Line

Security work makes prompting confidentiality-critical in a way most professions are not: your prompts contain indicators, infrastructure detail, vulnerability information, and sometimes evidence in active investigations. Know which tier you are typing into — consumer AI tools may retain and train on input, enterprise tiers contractually should not, and your organisation's AI use policy should name approved tools per data class. Habits that hold the line: sanitise before pasting — placeholder hostnames, masked IPs, and generic names preserve almost all the analytical value; never paste credentials, active-incident evidence, or unpatched vulnerability detail into anything unapproved; and remember prompt history is a record. Assume everything typed persists somewhere, and write accordingly.

  • Know your tier: consumer tools may train on input; enterprise terms should forbid it contractually
  • Sanitise by default — placeholders and masked indicators keep the analytical value
  • Credentials, live-incident evidence, and unpatched vulnerability detail stay out of unapproved tools
  • Prompt history persists: treat every prompt as a written record

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.