AI-Powered Agents for Bug Bounty Hunting and Offensive Security
Bug-Bounty-Agents is a collection of specialized AI agent prompts designed for bug bounty hunting, penetration testing, and offensive security workflows. These agents are structured as drop-in personas for large language models (LLMs) such as Claude Code, GitHub Copilot Chat, Cursor, and other agent-capable LLM clients. The project eliminates the need for frameworks or dependencies, focusing instead on disciplined prompts that transform generic LLMs into specialized tools with strict scope enforcement.
The repository contains 43 distinct agents, organized across six engagement phases: reconnaissance, web/API/application testing, infrastructure/cloud/network assessments, exploitation and post-exploitation, specialized/adversarial tasks, and defense/reporting/orchestration.
The project emphasizes that these are prompts, not scanners. They guide LLMs to act as specialists in specific domains, such as vulnerability hunting, exploit chaining, or reporting, while maintaining control over the engagement scope and methodology. Users retain full responsibility for driving the engagement and interpreting results.
How It Works
Bug-Bounty-Agents operates by providing pre-defined, production-ready agent personas that can be integrated into compatible LLM clients. Each agent is a self-contained Markdown file containing instructions, scope constraints, and methodologies tailored to a specific offensive security task. These files are designed to be portable across supported platforms, including Claude Code, GitHub Copilot Chat, Cursor, and generic chat interfaces like ChatGPT or Gemini.
The installation process is streamlined through a shell script (install.sh), which auto-detects the user’s LLM client environment and deploys the agents to the appropriate directories. For example, agents for Claude Code are placed in ~/.claude/agents/ (global) or .claude/agents/ (per-project), while Copilot Chat agents are converted to .chatmode.md files in the user’s VS Code prompts directory. Cursor agents are installed in .cursor/rules/, and generic clients require manual copy-pasting of the Markdown content into their system prompt fields.
Well-designed agents will prompt for clarifications, adhere strictly to the provided scope, generate reproducible proofs of concept (PoCs), and produce triage-ready findings with severity and impact assessments.
The project supports Burp Suite’s MCP Server integration, allowing LLM clients to interact directly with Burp Suite’s proxy, Repeater, Intruder, and site maps. For example, pairing the web-hunter or api-security agents with Burp MCP allows for dynamic testing and exploitation workflows.
Use Cases
The agents are categorized as follows:
- Reconnaissance & Intelligence: Agents such as recon-advisor, osint-collector, and subdomain-takeover focus on surface enumeration, asset discovery, and validation of dangling DNS records. The threat-modeler agent applies STRIDE or attack-surface modeling techniques to identify potential vulnerabilities early in the engagement.
- Web, API & Application: This category includes agents like web-hunter, api-security, and graphql-hunter, which specialize in testing web applications, REST/GraphQL APIs, and business logic flaws.
- The red-team-operator agent supports long-haul operations, including command-and-control (C2) and operational security (OPSEC).
- The swarm-orchestrator agent coordinates multiple agents in parallel, while scope-guard enforces hard scope boundaries during long-running engagements.
Users can combine agents in workflows to simulate full offensive security engagements. The swarm-orchestrator agent can manage parallel execution of these phases, while scope-guard ensures adherence to defined boundaries.
Key Features & Capabilities
Bug-Bounty-Agents prioritizes simplicity, portability, and specialization without introducing dependencies or frameworks.
- No Dependencies: Agents are self-contained Markdown files that require only an LLM client capable of processing system prompts or instruction files. This design ensures compatibility with a wide range of tools, from Claude Code to generic chat interfaces.
- Strict Scope Enforcement: Each agent includes built-in mechanisms to prevent drift from the defined target and scope.
- Reproducible Outputs: Agents are designed to produce reproducible PoCs and triage-ready findings, including severity and impact assessments. This ensures consistency and reliability in offensive security testing.
- Integration with Burp Suite: The project supports Burp Suite’s MCP Server, enabling LLM clients to interact with Burp’s proxy, Repeater, Intruder, and site maps.
- Modular Workflows: Agents can be combined in workflows to simulate full engagements, from reconnaissance to reporting.
- This breadth ensures comprehensive coverage of offensive security tasks.
- Open-Source & Permissive Licensing: Bug-Bounty-Agents is released under the MIT License, allowing for unrestricted use, modification, and distribution in both personal and commercial contexts.
The project also includes a disclaimer section, emphasizing that agents are tools to assist in offensive security testing and that users remain responsible for their actions. The repository provides examples of end-to-end engagements, such as web-bug-bounty.md, which demonstrates a full workflow from reconnaissance to report generation, culminating in a HackerOne submission.