Quick Answer: An AI jailbreak plugin is a third-party add-on that injects instructions to get an assistant to set aside its built-in safety rules. The "富江破甲" plugin for DeepSeek Harness, listed on GitHub, combines a prompt override with a skill library, an exploit knowledge base, and automated pentest workflows. Before you install anything like it, check what it writes to disk, what it overrides, and whether you have written authorization for every target.
A GitHub project called fujiang-pojia (富江破甲) calls itself a DeepSeek Harness plugin. Its description lists a trigger word, "direct answers to everything," a skill library, a CVE/PoC knowledge base, vulnerability report tracking, and "fully automated authorized penetration testing." Two phrases deserve a closer look: "破甲" literally means "break armor," and "fully automated" means the tool keeps choosing the next step. This article explains what a plugin like this actually changes, and how to evaluate it before it touches your machine or your targets.
What an AI jailbreak plugin actually changes
The mechanism matters more than the marketing. According to the project's README, the plugin edits the system prompt. When injection is enabled, it replaces the official harness:identity opening with an empty segment. That is the instruction layer that shapes how the model behaves, so this is not a cosmetic change.
Then add the agent loop. The README describes automatic continuation of up to 8 rounds, plus skills that the model can discover and load on its own. The cause-and-effect chain is simple: a prompt that discourages caution, combined with an agent that keeps going without a human checkpoint, means an early mistake compounds instead of stopping. A single bad assumption about a target can turn into eight rounds of actions before anyone reads the log.
Here's the part most people miss. The plugin's value proposition is removing friction, and friction is often the safety control. Refusals and confirmation prompts are annoying, but they're also where a human notices that a target looks wrong.
Why "authorized" in the description doesn't do the work
The word "authorized" appears in the project description, and a disclaimer is attached. Neither changes what the software can do. Authorization is a fact about an engagement: a signed statement of work, a defined scope, targets you own or have written permission to test. A prompt has no way to check any of that. The model cannot read your contract.
This is the counter-intuitive point. The more autonomous and less refusal-prone a tool is, the more all the risk sits with the operator. A disclaimer protects the author. It does nothing for the system you accidentally scanned because a fingerprint match looked plausible.
Legal exposure is real and varies by jurisdiction. China's Criminal Law Article 285 addresses illegal intrusion into computer information systems. The US Computer Fraud and Abuse Act and the UK Computer Misuse Act 1990 cover comparable conduct. Scope and written consent are what separate authorized testing from an offense, and an automated loop makes it easy to drift outside scope without noticing.
The knowledge base problem: public PoC is not a verified finding
The README is candid here. Knowledge entries are described as "对公开资料的整理" (organized from public material), and the metadata includes verified: false. That's the right label, but it's easy to lose when a tool presents a ranked list of exploits.
Consider Log4Shell (CVE-2021-44228). Public proof-of-concept code is abundant, but whether a given payload is safe or even effective depends on the Log4j version, JVM settings, network egress, and whether the target is a production system. A PoC that works cleanly in a lab can trigger crashes, log floods, or alerts in production. The matching logic also reflects this uncertainty. The README's weights (port 40, product 30, CVE 20, keyword 20, severity 5) form a heuristic ranking. A match on "port 8080, Tomcat" is a guess that needs confirmation, not a diagnosis.
| Source type | Typical verification | Main risk | Better use |
|---|---|---|---|
| Vendor advisories (e.g., Microsoft MSRC, Apache security pages) | Vendor-confirmed affected versions and fixes | Advisory may lag real-world exposure | Confirming version ranges before any testing |
| Maintained scanner templates (e.g., Nuclei community templates) | Peer-reviewed by a community with issue tracking | Templates can be outdated or false-positive prone | Detection with rate limits and manual confirmation |
| Aggregated or AI-generated knowledge entries | Often unverified, sources mixed | Wrong version ranges, untested payloads | Brainstorming candidates, never execution |
Most teams use the bottom row only as a starting list. Here's where it goes wrong: the list looks authoritative because it's organized, and organization feels like verification.
Installation is part of the attack surface
The install section of the README is the most useful part to read critically. It asks users to run a PowerShell installer, and it includes instructions for clearing the Windows "Mark of the Web" when Windows blocks the script. Mark of the Web is the Zone.Identifier marker Windows attaches to downloaded files. Clearing it is a legitimate step for scripts you have verified, but it's also exactly what a careless user does to get past a warning they didn't understand.
The README suggests checking the file hash with Get-FileHash. That's a good instinct, but understand its limit. A matching SHA256 proves the file is identical to what the publisher posted. It does not prove the publisher is trustworthy or that the code is safe. Compare the hash against a source you didn't get from the same page where you downloaded the file.
The README also states that the current version has checked file structure and install logic but has not completed a full runtime verification of every feature. That's an honest admission, and it means you're the one doing the testing. It also notes that prompt-custom.json may contain personal settings, which is a reminder that config files travel with the plugin.
Use this checklist before installing any AI agent plugin:
- Read the install script and every file it copies. Note any path outside the plugin's own directory.
- List the config files it overwrites, and back up your originals first.
- Test in an isolated VM or a separate, throwaway profile with no production credentials.
- Verify hashes against an independent source, not the download page itself.
- Write down the authorized scope for each target before the agent sees it.
Red flags are worth memorizing: trigger-word activation that changes behavior silently, framing that promises no refusals, unattended loops with no approval step, bundled exploit lists without verification status, and install instructions that ask you to override OS warnings.
Where AI belongs in a pentest, and where it doesn't
AI assistance has real value in security work, and it's not all-or-nothing. Practitioners commonly use models to summarize scan output, draft notes, map a fingerprint to candidate CVEs for a human to check, and format report sections. These tasks don't change the target.
The line sits at actions that touch a live system. Any request that sends traffic to a host should pass through a person who confirmed the scope. That's also where prompt injection becomes a concrete risk. The OWASP Top 10 for LLM Applications lists prompt injection as LLM01. An agent that reads web content, logs, or scan results can be steered by instructions hidden in that content. A jailbreak plugin deliberately weakens the guardrails that would otherwise catch this.
Most people stop reading at the feature list. Don't. The question that matters is not "what can this plugin do?" but "what happens when it's wrong, and who notices?"
Frequently Asked Questions
What is an AI jailbreak plugin?
An AI jailbreak plugin is an add-on that modifies an assistant's system prompt or instruction layer so it ignores or reduces its safety behavior. Some pair this with tools or knowledge bases. Because it changes the model's rules rather than the model itself, it depends entirely on the host platform and the configuration you run.
Is an AI jailbreak plugin safe to install?
Not by default. Check the install path, the files it overwrites, and whether its features have been tested at runtime. Run it in an isolated profile with no production access. Treat any claim of "authorized" use as something you must document yourself.
Is it legal to use an AI pentest plugin?
Using one is legal only when you have written authorization for each target and stay within the agreed scope. Laws such as China's Criminal Law Article 285, the US Computer Fraud and Abuse Act, and the UK Computer Misuse Act apply regardless of whether an AI tool performed the actions.
Why does a public exploit sometimes fail or cause damage?
Public PoCs are written for specific versions and configurations. A payload that works on an unpatched lab build can crash a production service, trigger alerts, or do nothing at all on a different version. Confirm affected versions against vendor advisories before running anything.
Before you install any AI jailbreak plugin, read what it changes, test it somewhere disposable, and write down the scope it's allowed to touch. Pick one plugin you already run, list every file it writes and every prompt it overrides, and note what it would do if you weren't watching. If the answer surprises you, that's your signal to slow down.