A MIT-licensed prompt pack from mdpsec that turns an LLM agent into a phased, evidence-first bug bounty assistant for authorized web application research and responsible disclosure.
| Tool | mdpsec/bug-bounty-hunting-prompts — reusable prompt files for a structured, evidence-first web bug bounty hunting workflow |
| Category | LLM prompt engineering / security research methodology |
| Primary Use | Guiding an LLM agent through phased recon, sweep, validation, and triage of findings on authorized bug bounty targets |
| Safe Use | Authorized security research only: in-scope bug bounty program assets, bounded non-destructive proofs, responsible disclosure through the program's channel |
| Telemetry Note | As pure prompt files it emits no network traffic itself; observable footprint depends entirely on the tools and agent the operator wires in, and the workflow explicitly forbids leaving state changes behind |
Most attempts to use LLMs in offensive security collapse into unstructured chatter: the model hallucinates endpoints, forgets scope, and produces confident nonsense that wastes triage time. mdpsec/bug-bounty-hunting-prompts takes the opposite approach — it is a curated sequence of markdown prompt files, extracted from a real private hunting workflow, that impose phase discipline on an agent. The repository sits at a healthy 488 stars under the permissive MIT license, and its value is less in any single prompt than in the orchestration philosophy it encodes: hand the agent one phase at a time, preserve handoff artifacts, and force evidence before escalation.
It is worth being precise about what this is not. The README is unusually candid that the public release deliberately strips out the original author's helpers, browser fleet, proxy and email services, dashboard, proof-of-concept runtime, credentials, private operating rules, and reference corpus. The prompts are intentionally non-runnable as-is; every private integration is represented by a YOUR_* variable or angle-bracket placeholder. The companion ADAPTATION.md is the replacement index for all of those, and it is explicitly the first thing to read. This honesty matters for operators evaluating effort: you are adopting a methodology skeleton, not a turnkey product.
The phase sequence is the core artifact. It opens with phase-00-recon.md for a lightweight surface map, moves through phase-01-access.md for account setup and authenticated access, then phase-02-sweep.md for a broad endpoint and vulnerability-class sweep. From there the workflow splits: phase-04a-scope.md and phase-05a-scope.md inject scope-only context, while phase-04b-browser-walk.md performs an authenticated browser walk and phase-05b-critical.md shifts into critical-impact hunting. Resource hygiene is not an afterthought — phase-03-closeout.md validates the sweep handoff and releases resources, a detail that reflects a workflow built for long-running real operations rather than lab demos.
The back half of the sequence is where the evidence-first claim earns its keep. phase-06-triage.md handles initial triage and deduplication, phase-07-validate-escalate.md validates and escalates findings, and phase-08-verify-escalate.md layers on independent verification — a second pass that catches the single-agent failure mode of confirming its own mistakes. phase-08a-independent-escalation.md adds a read-only pre-gate escalation pass, and phase-09-triage-gate.md acts as an independent live triage gate before anything approaches reportable status. Finally, phase-10-self-duplicate-check.md performs same-program duplicate convergence, which anyone who has burned a submission on a duplicate will appreciate.
Mechanically, these are prompt source files, not an application or orchestration engine. The README describes the adaptation loop plainly: choose the phases matching your workflow, replace runtime variables and integration commands per ADAPTATION.md, supply current target scope and researcher-owned accounts at runtime, then feed the agent one phase at a time while preserving handoff artifacts. There is no daemon, no API wrapper, no pip install — you supply the model, the tooling, and the glue. That makes it tool-agnostic by construction, portable across whatever agent framework an operator prefers.
The modularity is more thoughtful than it first appears. The README notes that an unauthenticated workflow can simply omit account setup, browser leases, email, SMS, and authenticated walks while retaining the recon, sweep, triage, and reporting principles. That trimming matters in practice, because many programs and lab environments do not warrant a full identity-layer simulation. It also means blue teams and educators can adopt the methodology for purely defensive review — running the recon and sweep prompts against systems they own to see what surface an automated hunter would map.
The safety posture is written into the prompt layer itself, which is the right place for it. The instructions direct the operator to run only against authorized, in-scope assets, use bounded non-destructive proofs, and restore reversible test changes. The safety section goes further: do not target real users, do not automatically submit or publish findings, do not collect data in bulk, and do not leave state changes behind. Secrets hygiene is explicit — the README warns not to paste credentials, cookies, tokens, or API keys into the prompt files. This is a workflow designed to keep a human in the disclosure loop.
From a defender's perspective, this repository is also a useful threat-model artifact. Nothing here is an exploit; it is methodology. But the structure — automated surface mapping, authenticated browser walking, vulnerability-class sweeps, and dedup-aware escalation — mirrors exactly what sophisticated automated scanners do against exposed estates. Security teams running authorized internal exercises can use the phase list as a checklist for purple-team walkthroughs: if phase-00-recon-style surface mapping surprises you on your own assets, your asset inventory has a gap.
The maintainer sets expectations firmly: the pack is provided as-is, with no setup, integration, debugging, or usage support, and the private helpers, RAG corpus, and orchestration services will not be shared or reproduced. The status section frames it as a reference prompt pack expected to evolve as hunters adapt it to different agents, tools, and program rules. That is an honest sustainability posture for a methodology release — the artifact is meant to be forked and reshaped rather than consumed verbatim, and adaptation is the intended contribution model.
For the authorized professional, the appeal is leverage with guardrails. Structured prompt packs like this address the two chronic failure modes of LLM-assisted research — scope drift and unverifiable findings — by making scope injection (phase-04a, phase-05a) and independent verification (phase-08, phase-09) explicit, mandatory stages. Anyone building an agent-assisted hunting or assessment pipeline would do well to study the handoff-artifact pattern here even if they never run a single prompt as written. Combined with its MIT license and clean provenance, it is a low-cost, high-signal addition to an authorized testing toolbox.
mdpsec/bug-bounty-hunting-prompts.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.