Skip to content

Back to the catalog

matty69v/Bug-Bounty-Agents
github.comrepo

matty69v/Bug-Bounty-Agents

Bug-Bounty-Agents is a collection of 43 specialized LLM system-prompt personas (recon, web hunting, exploit chaining, reporting, scope-guard, etc.) that turn Claude Code, Copilot Chat, Cursor, or any agent-capable LLM into offensive-security specialists for bug bounty and pentest engagements. It ships an install script that auto-detects and deploys agents to the correct client directories, and documents pairing the agents with PortSwigger's Burp Suite MCP server so an LLM can drive live traffic through Burp while following the repo's methodology.

TL;DR Drop-in LLM agent prompts (no framework/deps) for structuring bug bounty engagements, pairable with Burp Suite via MCP.
Document

What it is

matty69v/Bug-Bounty-Agents is a repo of 43 markdown system-prompt personas for offensive-security work, installed via ./install.sh into Claude Code, Copilot Chat, Cursor, or pasted manually into ChatGPT/Gemini/API clients. The page adds the concrete install/uninstall commands, the per-client target directories, the engagement pipeline, and build steps for pairing with PortSwigger's Burp MCP server — none of which the summary specifies.

How it works

Each .md file is a focused persona. The installer auto-detects claude, code, and cursor on your PATH and copies files to the right location, renaming as needed (e.g. .chatmode.md for Copilot). Agents map to six engagement phases; the machine-readable index is AGENTS.md. A typical run chains:

  • recon-advisor → enumerate attack surface
  • web-hunter / api-security → find vulnerabilities
  • exploit-chainer → escalate impact
  • poc-validator → confirm and stabilize
  • report-generator → write the submission

Layer scope-guard on top of any agent to hold scope boundaries during long sessions; swarm-orchestrator or attack-planner coordinate a full engagement; red-team-operator and purple-team run side by side for purple-team work.

What to do

Install commands:

./install.sh                     # auto-detects your client(s)
./install.sh --target claude         # Claude Code (global)
./install.sh --target claude-local   # Claude Code (this project)
./install.sh --target copilot        # Copilot Chat (VS Code)
./install.sh --target cursor         # Cursor (this project)
./install.sh --target all            # everything detected
./install.sh --dry-run --target claude
./install.sh --uninstall --target claude

Client directories for manual install: Claude Code reads ~/.claude/agents/ (global) or .claude/agents/ (per-project); Copilot uses the VS Code User/prompts dir with a .chatmode.md rename; Cursor uses .cursor/rules/.

To wire up Burp Suite, build the PortSwigger MCP jar (requires Java and jar on PATH):

git clone https://github.com/PortSwigger/mcp-server.git
cd mcp-server
./gradlew embedProxyJar
# output: build/libs/burp-mcp-all.jar

Then in Burp: Extensions → Add → Extension Type Java → select the jar → open the MCP tab and tick Enabled. Load web-hunter, api-security, ssrf-hunter, or bizlogic-hunter alongside it so the agent can act through Burp.

Caveats

The page is explicit that these are prompts, not scanners: "They make an LLM act like a specialist; they do not bring their own tooling. You still drive the engagement." Zero dependencies means no bundled capability — behaviour depends entirely on the host LLM honouring the scope enforcement the prompts describe. Enabling MCP tools that can edit your config is gated on trusting the client.

Technologies named here

Techniques named here

  • Recon / attack surface enumeration
  • Web vulnerability hunting
  • API security testing
  • Exploit chaining to escalate impact
  • PoC validation and stabilization
  • Reporting/triage-ready findings with severity and impact
  • Auth bypass testing on API endpoints
  • IDOR testing on /api/v2/users endpoints
  • SSRF hunting
  • Business logic vulnerability hunting
  • Scope enforcement during engagements
  • Red team / purple team collaborative operations