wtf is llm-hunter?

LLM Hunter is a Claude Code plugin that organizes an AI coding agent into a structured team for authorized penetration testing and bug bounty work. Instead of letting one model try to do everything from scratch each time, it splits the work across multiple agent roles that each focus on one part of the process, such as reconnaissance, attacking, and creative problem solving. The README stresses that the only real guardrail is authorization. The engine will only test what a program's rules explicitly allow, checked against a rules file created for each engagement, and it never starts without a verified bug bounty program page or a signed pentest agreement. It does not filter out approaches just because they seem unlikely to work, since the README argues that self-censoring on relevance causes agents to miss real findings. A key feature is what the README calls the anti-give-up mechanism: when an approach comes back negative, a controller keeps retrying with genuinely new angles rather than accepting the first negative result as final. There is also a separate pool of agents that only sees the raw attack surface, without knowing what other agents already tried, so their ideas are not influenced by earlier failures. Anything useful an agent discovers during a run gets written up as a reusable skill, so the tool is meant to improve the more someone uses it. The system works with different language models, described in tiers like cheap, mid, and strong rather than by naming specific vendors, so it can adapt to whichever models are available. To use it, someone installs it as a Claude Code plugin, then runs a single slash command with the program's URL or scope. The pipeline handles a legitimacy check, defines scope, runs a recon and attack cycle, and always finishes with a report, even if nothing is found, before asking whether to keep going. It runs entirely on an existing Claude Code subscription rather than requiring a separate API key.