13 AI agent prompts for Security

Copy-paste, beginner-friendly prompts to help you triage alerts, review access and write security policies in plain English. Each one is based on a real-world example and works with ChatGPT, Claude or an AI agent. 6 are free below, and 7 more are in AgentPlaybook Pro.

Free Security prompts

Produce a first-pass malware reverse-engineering brief for human researchers. Sample / hash / sandbox report: [paste or attach]. Environment notes: [lab constraints]. Cover: (1) high-level behavior summary, (2) persistence and execution mechanisms, (3) network indicators and C2 patterns if present, (4) notable code or packing tricks, (5) similarities to known families if evidence supports it, (6) open questions and recommended next analyst steps, (7) IOCs in a clean table. Label confidence on every claim and never invent hashes or C2 hosts. Don't detonate samples outside the approved lab, don't block or hunt in production, and leave containment decisions to me.

I want to use an AI agent for these tasks: [list, e.g. reorder groceries, book flights, pay utility bills, update my store listings]. Accounts I have: [list sites and apps, no passwords]. My password manager: [name, or 'none']. 1) For each task, list the minimum accounts the agent actually needs and mark any it doesn't need. 2) Sort those accounts into three groups: OK to let the agent log in with my approval each time, OK only for viewing, and never (for example banking, email recovery, government ID). Explain each choice in one line. 3) For each allowed account, list the actions that must always ask me first, such as paying, sending messages, changing addresses or deleting anything, and suggest a spending cap. 4) Give me step-by-step setup instructions that never involve typing a password into the chat, using my password manager's agent or sharing features if it has them. 5) Give me a monthly 5-minute checklist to review what the agent did and remove access I no longer need. Only use what I listed; don't guess about features my apps may not have.

Act as my personal security auditor. Ask me one question at a time, in plain language, about: my main email account, my bank and payment accounts, my phone, how I store passwords, and any apps or AI agents I've connected to my email, calendar, files or bank. Never ask me for a password, code, PIN or account number, and tell me to stop if I start typing one. Wait for each answer before the next question. When you're done, give me a ranked fix list in three groups: Today (under 10 minutes each), This week, and This month. For each fix say why it matters for me specifically and the exact setting to look for. Always check these basics: auto-updates on, passkeys or an authenticator app instead of SMS codes, a password manager, which apps and agents can read my email, and a recovery email and phone I actually check.

I run an AI agent (or other apps) on a rented Linux server: [provider, e.g. Hetzner/DigitalOcean], [OS, e.g. Ubuntu 24.04]. Right now I reach it by [SSH on a public IP / a public web address]. Services on it: [list, e.g. agent dashboard on port 3000, n8n on 5678]. Help me take it off the public internet using Tailscale for my own access and, only if something must stay public, a Cloudflare Tunnel. Give me one step at a time with the exact command, what I should see if it worked, and wait for me to confirm before the next step. Order it so I can't lock myself out: install and test Tailscale access first, then close public ports in the firewall last, and tell me how to get back in through my provider's web console if something goes wrong. Finish with a short checklist to confirm nothing is still reachable from the public internet. Don't ask me for passwords or keys; tell me where to enter them myself.

Every morning at [TIME], act as the patch manager for the computer my AI agents run on. 1) Collect the last 24 hours of security advisories (new CVEs, zero-days, actively exploited bugs) from [SOURCES, e.g. official vendor advisories, the CISA Known Exploited Vulnerabilities list, trusted security accounts on X]. 2) Read-only first: list what's actually installed on this machine (OS packages, language runtimes like Node, Python and Go, pip/npm packages, containers) and match each advisory to installed software. Ignore advisories for software that isn't installed and say so. 3) Apply on your own ONLY patch-level updates from official sources, and verify each one against its official checksum or signature before installing. 4) Do NOT apply major-version upgrades, changes to code repositories, anything touching secrets or keys, or anything that needs a restart of [CRITICAL SERVICES]. List those for my approval with the risk of waiting. 5) After updating, check that [KEY SERVICES/AGENTS] still run. Report: advisories reviewed, what applied to me, what you patched (old to new version, checksum verified), what's waiting for my yes, and any sign of compromise you noticed.

I'm doing authorized security testing. The program is [bug bounty program or my own app], and its scope rules are: [paste in-scope targets, out-of-scope items and testing rules]. Work in this order and stop to ask me before any step that could change data or send many requests. 1) Restate the scope and list anything you will NOT touch. 2) Map the attack surface from what I give you (URLs, API docs, JavaScript files, features): list features worth testing and why. 3) For each, name the likely vulnerability types (for example access-control mistakes, injection, SSRF, auth bypass) and a safe, low-impact test. 4) When I paste a result, decide: real issue, false positive or needs more proof, and say what evidence is missing. 5) For confirmed issues, write a report: title, severity with reasoning, affected endpoint, steps to reproduce, impact, evidence and a suggested fix. Never test outside scope, never access other users' real data beyond what's needed to prove the issue, and never run denial-of-service tests.

More in AgentPlaybook Pro

• Threat investigation plan and case summaryGather detections, indicators, and context into an investigation plan, then draft a case summary with recommended response actions for an analyst to approve—like Sophos MDR agents built with OpenAI Daybreak that cut average response time from about 38 minutes to 89 seconds.
• Backtest a detection rule before you ship itBefore changing a SIEM or detection rule, have an agent replay the proposed change against months of stored alert evidence and summarize what would have fired or gone quiet, like Block's security team replaying a detection change against seven months of telemetry in under an hour with OpenAI Daybreak.
• Run a vendor security risk assessment in minutesTurn a weeks-long vendor security review into a fast first-pass assessment your team still signs off—like League's Claude skill that finished vendor security risk assessments in about 15 minutes, with 49 of the last 53 validated as safe before human approval.
• Bait scammers for fraud intelligence before customers get hurtDeploy undercover AI agents that answer scam calls and messages, keep fraudsters talking, and extract bank accounts, wallets, malicious sites, and impersonated brands for your fraud team—before a real customer is reached. Apate's agents have run millions of these conversations for banks and telcos like Commonwealth Bank and TPG.
• Run a structured security audit of your code with a coding agentTurn your coding agent into a careful security auditor: it maps how your app works and where untrusted input comes in, hunts for weaknesses area by area, then has a separate check try to disprove each finding before anything lands in the report. Cloudflare open-sourced this as its security-audit skill, the starting point for its own vulnerability-discovery system, with six phases from reconnaissance to verified findings and a plain report.
• Check your own Android phone for spyware and risky apps: connect it to your laptop and let Claude Code run a read-only security auditA viral X thread this week showed Claude Code being used for personal security checks, and Raytar shared on X (Oct 2026) the prompt he uses: plug your Android into a laptop, turn on USB debugging, open Claude Code and have it act as a phone security auditor over adb (Android's built-in developer connection), on your own phone or one you have permission to check, with strict instructions not to delete, disable or change anything. This version keeps the audit read-only and turns the findings into a plain-language report you act on yourself.
• Before you let Claude Code work without asking, set guardrails: write deny rules for anything risky, add a hook that blocks dangerous commands before they run, and turn on sandboxingAs coding agents move toward running steps without asking first, people are setting limits up front. A post on X by @Cuiszt_Cui (Oct 23, 2026) laid out 6 steps for Claude Code permissions, hooks and sandboxing, noting that in Claude Code a deny rule beats any allow rule and a hook can block a command before it runs. This entry turns that into a beginner-friendly setup: decide what the agent may never do, write it down as rules, and test that the blocks actually work before giving the agent more freedom.

See Pro pricing →

Guides that include Security prompts

Other categories