Doing authorized security work in the terminal with free abliterated models

You paste a stack trace from your own app and ask how the bug could be exploited, so you can fix it properly. The model says no.

I hit this a lot. Most coding models get nervous around anything that looks like an attack, even when the system is yours.

This post shows how I do that kind of work in the terminal with FreePi, a free AI coding agent. It covers what abliterated models are, how to pick one, and two real workflows: auditing your own repo and analyzing software you're allowed to take apart. Everything here is for authorized work only.

Try it: run npx free-pi-cli@latest in a terminal. No card, no API key.

Give this to your agent: freepi.ai/llms.txt

The ground rules first

This only covers work you're allowed to do:

• Systems and code you own.
• Employer or client systems where you have written permission to test.
• Bug bounty targets, inside the program's stated scope.
• CTFs and lab boxes built to be attacked.
• Software you own or may lawfully analyze.

A different model doesn't change what you're allowed to do. Laws like the US Computer Fraud and Abuse Act still apply to you. If you're not sure you're authorized, find out before you start.

What "abliterated" means

An abliterated model is an open-weight model with its refusal behavior removed. It will help with the lawful parts of security work that general models often decline, like writing a proof of concept for a bug in your own handler.

In FreePi I call these abliterated models for defense. They're opt-in. You only get one if you pick it.

They are not smarter. They just refuse less. You still check everything they say.

Setup

You need Node.js 22.19 or newer. Then:

npx free-pi-cli@latest

You may start as a guest. The defense models need a GitHub sign-in, though. The server refuses them for guest sessions. Type /login to sign in with GitHub. If you pick a defense model as a guest, it opens the sign-in for you.

Run it from the project directory you want to work in. The agent reads, searches and edits files there.

Pick a model for defense

Inside the agent, type:

/model

The picker is built from the server's model list when the CLI starts. The abliterated models for defense sit near the top, labeled free-pi-defense. Today that list is:

• Abliterated Large Flash
• Abliterated Large V2
• DeepSeek V4.1 Flash Abliterated
• Qwen 3.6 Plus Uncensored
• Gemma 4 Uncensored
• GLM 4.7 Flash Heretic
• Uncensored 1.2

The list can change without a new CLI release. Whatever you see in /model is current.

My habit: start on the default model. Switch to a defense model only when the general one refuses the lawful part of the job. Then switch back.

Workflow 1: audit a repo you own

FreePi bundles a security audit skill, based on Cloudflare's open-source security-audit skill. A skill is a set of instructions the agent loads for one kind of job. You can ask for it in plain words, or force it by name:

/skill:freepi-security-audit security audit this codebase

It works in two modes.

Guidance mode is for a focused question, like "is this upload handler safe?" It uses the relevant parts of the method and writes no report files.

Full audit mode runs when you ask for a full audit, a pen test of the codebase, or a report. It works in phases: map the architecture and trust boundaries, hunt for weaknesses, try to disprove each candidate, then write the results.

A full run writes these files:

• architecture.md and coverage-ledger.json
• findings.json, with each record marked confirmed, needs_validation or rejected
• REPORT.md, FINDINGS-DETAIL.md and NEEDS-VALIDATION.md

By default they go outside your repo, in ~/security-audit-skill/<repo-name>/run-<N>. The skill is told never to edit your source.

Two things to know:

1. It runs as one agent. The same agent that finds an issue also tries to disprove it. So you are the second reviewer. Follow each confirmed finding into the code yourself.
2. It validates by reading code. Before it runs anything from the target, it is told to ask you and show the exact command.

I find it helps to state scope at the top of the session:

This is my own repo. I own the code and the
staging server. Review src/api for auth bugs.
Read only. Ask before running any command.

Workflow 2: look inside software you may analyze

Sometimes the question is about a binary, not source. Say a CLI tool you're licensed to use, or an Electron app you shipped years ago and lost the source for.

Type:

/reverse-engineer-anything

This loads REA (Reverse Engineer Anything) for the current session only. It gives the agent reverse-engineering tools for native binaries, JS and Electron apps, APKs, .NET assemblies and websites.

A few practical notes:

• The first run downloads the tools, so it can take a minute.
• It needs Node 22.19+, 24.11+ or 26+. If your Node is older, it tells you.
• Deep native analysis needs your own Hopper, Ghidra or IDA install.

Then ask in plain words. "List the modules this Go binary was built from, with versions." Or "show me this Electron app's fuses and IPC handlers."

Keep it safe on your machine

The agent has real tools: read, write, edit, bash, grep, find and ls. In a normal session FreePi has no OS sandbox. bash runs on your machine.

So:

• Work in a container or a throwaway VM when you test anything risky.
• Never point it at production. Use a local or staging copy.
• Keep real credentials out of that environment.
• Keep your work under version control.
• Press Esc to interrupt the agent if it heads somewhere you didn't ask for.

And reproduce every finding yourself before you report it. A model can be confidently wrong, abliterated or not.

Wrapping up

Abliterated models for defense won't do your security work for you. They stop the refusals that get in the way of lawful work on your own systems. Pair one with the audit skill for code, and with /reverse-engineer-anything for binaries. Set your scope, work in a sandbox you control, and check every result.

Try it free with npx free-pi-cli@latest, and read more at www.freepi.ai.