Claude Coding

6 Prompts to Audit Your Own Self-Hosted AI Setup

A cryptomining botnet spent six months hunting LiteLLM and Ollama servers left open to the internet, hijacking their GPUs and taking their API keys. Most victims did not know they were exposed. These six prompts walk an AI model through auditing infrastructure you own: your config files, your firewall rules, your logs, your keys. Paste your own files into them. Every one is scoped to systems you control, and none of them will help you look at anyone else.

⌨️ 6 prompts 🕐 Updated Oct 13, 2026

💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.

The Exposure Check

The single most common cause of compromise is a service bound to all interfaces by someone who assumed it was local. Asking only for the inventory, with no fixes, stops the model skipping straight to advice.

Here is my docker-compose file and my service config. List every port this exposes and state for each one whether it is reachable from the public internet, from the local network only, or from localhost only. Flag anything bound to 0.0.0.0. Do not suggest improvements yet. Just tell me what is open. [PASTE FILES]

The Firewall Rule Writer

The undo clause matters. Most firewall advice omits it, and locking yourself out of a remote machine is the reason people leave ports open in the first place.

I need to restrict access to a service on port [PORT] so it is reachable only from [TRUSTED RANGE] and from localhost. My server runs [OS AND FIREWALL]. Give me the exact commands, then tell me how to verify the rule is working from outside the machine, and how to undo it if I lock myself out.

The Log Hunt

Asking what looks normal as well as what looks wrong gives you a model that reasons rather than one that pattern matches a scary answer out of ordinary traffic.

Here is a sample of my outbound connection logs. I am looking for signs of a cryptominer: persistent connections to unfamiliar addresses, regular beaconing at fixed intervals, and traffic to mining pool ports. Tell me what looks normal, what looks unexplained, and what you would need to see more of to decide. [PASTE LOGS]

The Key Blast Radius

Rotation order is the part people get wrong under pressure. Rotating a key that another service depends on before updating that service turns an incident into an outage.

My server at [DESCRIPTION] may have been compromised. Here are the environment variables and config files it had access to, with the secret values removed. For each credential, tell me what an attacker could do with it, in order of damage, and what the correct rotation order is so I do not lock out my own services mid-way.

The Update Inventory

AI tooling sits outside most patch cycles because it arrived after the cycle was designed. Asking for the smallest process you will actually follow beats a thorough one you will abandon.

Here is the list of AI and developer tools running on my server with their versions. For each one, tell me whether it is current, when it was last updated, and whether it belongs in a regular patch cycle. Then give me the smallest reasonable update process I could actually keep up with as one person. [PASTE LIST]

The Honest Second Opinion

The most dangerous state after a security review is believing it is finished. Explicitly asking a model to argue against your work surfaces the gaps that a confirming answer hides.

I have just hardened my self-hosted AI server as follows. [DESCRIBE WHAT YOU DID] Argue against me. What did I miss, what did I do that gives false confidence, and what would still be exposed if my assumptions are wrong. Do not reassure me.