hackerone

Files (7)

• SKILL.md
• INTEGRATION_GUIDE.md · reference/
• sensitive-data-tracking.md · reference/
• __init__.py · tools/
• csv_parser.py · tools/
• report_validator.py · tools/
• sensitive_data_tracker.py · tools/

SKILL.md (first lines)

# HackerOne Bug Bounty

Automates: scope parsing → parallel testing per asset → **authoritative finding validation** → submission reports.

## Quick start

1. Input: HackerOne program URL or scope CSV.
2. Parse scope and program guidelines.
3. Spawn one coordinator per eligible asset (parallel).
4. Each coordinator runs the standard engagement flow (see `skills/coordination/SKILL.md`).
5. Run the **`validate-findings` workflow per asset** (authoritative submission gate). Generate HackerOne reports from the validated set ONLY — never submit a finding that is not `VALID`/`REPAIRED`.

## Scope CSV format

Expected columns:
- `identifier` — asset URL/domain.
- `asset_type` — URL, WILDCARD, API, CIDR.
- `eligible_for_submission` — must be `true`.
- `max_severity` — critical / high / medium / low.
- `instruction` — asset-specific notes.

Parse with `skills/hackerone/tools/csv_parser.py`. Filter for `eligible_for_submission=true`.

## Agent deployment

One coordinator per asset, spawned in parallel:

```python
coordinator_role = Read("skills/coordination/SKILL.md")
Agent(prompt=f"{coordinator_role}\n\nTARGET: {asset_url}\nSCOPE: {program_guidelines}\nOUTPUT_DIR: ...",
      run_in_background=True)
```

10 assets → 10 parallel coordinators (~2-4 h vs 20-40 h sequential). Each coordinator follows `skills/coordination/SKILL.md` and `reference/role-matrix.md`.

## Finding validation (authoritative submission gate)

Every finding requires `poc.py` (executable exploit), `poc_output.txt` (timestamped execution proof), manual repro steps, and evidence (screenshots / HTTP captures / video). This is the input the validator consumes.

**After each asset's coordinator returns, the orchestrator runs the `validate-findings` workflow for that asset and submits ONLY findings it marks `VALID` or `REPAIRED`:**

```python
v = Workflow(name="validate-findings", args={
    "output_dir": asset_output_dir,   # the coordinator's OUTPUT_DIR
    "target": asset_url,
    "business_tier": "revenue",       # production/external bug-bounty scope; else "unknown"
    "votes": 3,                       # adversarial refuters per finding (bounty = stricter)
    "repair": True,                   # regenerate broken/absent PoCs so each finding has a runnable evidence script
    "strict": True,                   # one failed gate => REJECTED
})
# v.counts -> {total, valid, repaired, rejected}; v.validated[] / v.rejected[]
# Verdicts at {asset_output_dir}/artifacts/validated/{id}.json (+ false-positives/).
# Asset report at {asset_output_dir}/reports/validation-report.md.
```

`validate-findings` is the authoritative gate because it directly preempts the most common HackerOne rejections: it verifies each CVE against **NVD + a recomputed CVSS base score (from the vector) + CISA KEV + the vendor advisory**, **runs (and repairs) every PoC** until it emits the evidence that proves the issue, **recomputes the risk/severity**, corroborates every claim against raw evidence, and kills false positives via **adversarial refutation**.

The orchestrator is the only layer that runs this workflow (it is top-level; the workflow is one level below — legal nesting). To avoid double work, coordinators may treat their **interleaved per-finding validators** (run the instant each candidate is materialized) as a provisional self-check; the standalone `validate-findings` verdict is authoritative for submission. A `REJECTED` finding is a false positive — it never enters a submission, an appendix, or a count; its `artifacts/false-positives/{id}.json` is the sole record. Likewise a coordinator-dropped candidate (uncured DEMOTED, in `artifacts/dropped/`) is never submitted.

The HackerOne PoC contract is a superset of the standard finding contract (`skills/coordination/reference/validator-role.md`).

## Submission report format

Build each report from a validated finding only. Required sections per HackerOne standard:
1. Summary (2-3 sentences).
2. Severity (CVSS + business impact; v4.0 primary, v3.1 → v3.0 → v2.0 fallback) — use the score `validate-findings` recomputed from the vector and the risk bucket it assigned, not a hand-typed number.
3. Steps to Reproduce (numbered, clear) — mirror the validated `verification-script.py`.
4. Visual Evidence (from `evidence/validation/`).
5. Impact (realistic attack scenario).
6. Remediation (actionable fixes).

Validate report *format* with `skills/hackerone/tools/report_validator.py` (complements the finding validation).

## Output structure

Standard `OUTPUT_DIR` (`skills/coordination/reference/output-discipline.md`) plus a per-asset `reports/submissions/` containing the platform-ready markdown. `validate-findings` writes the verdicts under `artifacts/validated|false-positives/` and the asset `reports/validation-report.md`.

```
{OUTPUT_DIR}/
├── findings/
│   └── finding-NNN/evidence/validation/   # validate-findings proof packages
├── reports/
│   ├── submissions/                        # built ONLY from artifacts/validated/
│   │   ├── H1_CRITICAL_001.md
│   │   └── H1_HIGH_001.md
│   ├── validation-report.md                # validate-findings asset report
│   └── SUBMISSION_GUIDE.md
├── recon/
├── logs/
└── artifacts/
    ├── validated/          # {id}.json — submit these
    └── false-positives/    # {id}.json — never submit these
```

## Program selection

**High-value:** new programs (< 30 days), fast response (< 24 h), high bounties, large attack surface. **Avoid:** slow response (> 1 week), low bounties, restrictive scope.

## Submission checklist

- [ ] **`validate-findings` verdict = `VALID` or `REPAIRED`** (`artifacts/validated/{id}.json` exists). Never submit a `REJECTED` finding.
- [ ] CVSS **recomputed from the vector** (v4.0 primary; v3.1 → v3.0 → v2.0 fallback) and matching NVD when a CVE applies (`evidence/validation/cve-verification.md`).
- [ ] Working PoC with `poc_output.txt` (the validator re-ran/repaired it and confirmed the evidence token).
- [ ] Step-by-step reproduction.
- [ ] Visual evidence.
- [ ] Realistic impact (risk score/bucket from `evidence/validation/risk-assessment.md`).
- [ ] Remediation guidance.
- [ ] Sensitive data sanitized.
- [ ] Asset is `eligible_for_submission=true`.

## Common rejections (preempt)

| Rejection | Prevention |
|-----------|------------|
| Out of Scope | Verify `eligible_for_submission=true` and asset-type match |
| Cannot Reproduce | `validate-findings` re-runs (and repairs) the PoC 3× and requires a vuln-class evidence token — only reproducible findings pass the gate |
| Duplicate | Search disclosed reports before submission; submit quickly |
| Insufficient Impact | Document realistic attack scenario; `validate-findings` recomputes risk/severity so the rating is defensible |
| Incorrect severity / CVSS | `validate-findings` recomputes the base score from the vector and cross-checks NVD + KEV; any mismatch is flagged before you submit |

From transilienceai/communitytools, MIT.

More from transilienceai/communitytools