agentleFS
ivRodriguezCAivRodriguezCA/MobHunt/.claude/skills/mobile-hunting/SKILL.md

mobile-hunting

Skill37 starsChanged 3 months agoMITView on GitHub

Load when running or coordinating a full mobile hunt end to end, covering pipeline phase discipline and time budgets, hunter spawning and skip rules, confidence calibration, agentic loop termination, and dead end documentation.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 12 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈12

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    All vulnerability classes covered with documented dead ends → stop

    line 57
  2. Distinctive:

    All hunters report clean → stop with "clean" result

    line 59
  3. Distinctive:

    No URL schemes registered → skip IPC hunter

    line 86
  4. Distinctive:

    No native libraries → skip native hunter

    line 87
  5. Distinctive:

    No WebViews in source → skip WebView hunter

    line 88
  6. Distinctive:

    No embedded frameworks → skip SDK hunter

    line 89
  7. Distinctive:

    Use Haiku for scope (structured extraction, not reasoning)

    line 134
  8. Distinctive:

    Use Sonnet for validation and chaining (structured decision-making)

    line 135
  9. Distinctive:

    When a vulnerability class yields nothing, document it:

    line 144
  10. Distinctive:

    Before any finding advances to validation, pre-check against the never-submit list:

    line 166
  11. Distinctive:

    At least one chain attempt: if standalone findings are weak

    line 188
  12. Distinctive:

    No dead-end documentation (forcing redundant re-hunts)

    line 194

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Mobile Hunting: Discipline & Methodology

This skill defines the overall mobile vulnerability hunting methodology for MobHunt. It is loaded by /hunt and used by the orchestrator to coordinate the full pipeline.


Hunting Philosophy

Mobile bug bounty hunting is constrained by three realities:

  1. Static analysis is primary. Unlike web targets, you cannot freely probe mobile APIs. Binary analysis, decompilation, and configuration review are your main tools.
  2. Most findings are informational. The gap between "misconfiguration" and "exploitable vulnerability" is where bounties live. Bridge it with concrete attack scenarios.
  3. Chains beat singles. A medium-severity URL scheme issue chained with a WebView JS bridge becomes a critical RCE. Always think in chains.

Pipeline Methodology

Phase Discipline

SCOPE → ACQUIRE → RECON → HUNT → CHAIN → VALIDATE → REPORT
                                    ↑                    |
                                    └── agentic loop ────┘

Each phase has a clear deliverable. Do not skip phases or combine them:

PhaseDeliverableTime Budget
SCOPEscope.json with assets, boundaries, exclusions2 min
ACQUIREExtracted + decrypted app in workspace/5 min
RECONAttack surface map in recon/10 min
HUNTRaw findings in findings/30 min (5 hunters × 6 min each)
CHAINChain proposals in findings/chains/10 min
VALIDATEValidated/killed findings10 min
REPORTSubmission-ready reports in reports/10 min per finding

Agentic Loop Strategy

The loop re-hunts when validated_findings < 2 and loop_iteration < max_iterations:

Re-hunt priorities (in order):

  1. Vulnerability classes not yet covered by any hunter
  2. Areas where tools flagged signals but no hunter investigated
  3. Cross-platform patterns (if both iOS and Android are in scope)
  4. Deeper binary analysis on functions classified as security-relevant
  5. SDK-specific research for flagged but unversioned SDKs

Loop termination signals:

  • 2+ validated findings → stop (enough for submission)
  • All vulnerability classes covered with documented dead ends → stop
  • Max iterations reached → stop with whatever was found
  • All hunters report clean → stop with "clean" result

Hunter Spawning Strategy

Instance Assignment

The orchestrator spawns up to 5 hunters per platform. Assignment is based on recon signal density:

iOS hunters (ordered by typical yield):

  1. IPC & Entry Points: URL schemes, universal links, extensions
  2. Binary & Secrets: hardcoded keys, crypto misuse
  3. Data Storage: Keychain, UserDefaults, file protection
  4. Network Security: ATS, cert pinning, endpoints
  5. Third-Party SDKs: framework CVEs, debug frameworks

Android hunters (ordered by typical yield):

  1. Component Exposure: exported activities/services/providers
  2. Intent & Deep Links: deep link hijacking, intent redirection
  3. Data Storage: SharedPrefs, providers, SQL injection
  4. Network & WebView: WebView JS bridges, file access, cleartext
  5. Native & Secrets: JNI analysis, hardcoded secrets in .so

Skip Optimization

Skip a hunter instance if recon shows no surface:

  • No URL schemes registered → skip IPC hunter
  • No native libraries → skip native hunter
  • No WebViews in source → skip WebView hunter
  • No embedded frameworks → skip SDK hunter

Reallocate to a second pass on high-signal areas instead.

Focus Narrowing

Each hunter gets a narrow objective. Examples:

  • "Check if the myapp-auth:// URL scheme is used for OAuth callbacks and can be hijacked"
  • "Check the exported FileProviderActivity for intent injection or path traversal"
  • "Determine which Firebase project the API key AIzaSy... points at and whether that project is a listed in-scope asset"

Broad objectives ("find vulnerabilities") produce noise. Narrow objectives produce findings.

Authorization boundary: when an objective would send a request to a live service, narrow it further before spawning the hunter. Name the specific asset, confirm it appears in the program's stated scope, and confirm the program has given written permission to test it when the service is hosted by a third party (Firebase, Google Cloud, AWS, and similar); the program owns that account, not the researcher. Without both, the objective is a static one: establish what the credential or endpoint is documented to grant and hand the verification to the program.


Confidence Calibration

LevelCriteriaAction
HIGHClear vulnerability with demonstrable PoC. You could write the report now.→ VALIDATE → REPORT
MEDIUMLikely vulnerability, but needs runtime confirmation or chain partner.→ VALIDATE (may become CHAIN_REQUIRED)
LOWSuspicious pattern, may be false positive or dead code.→ Document as dead end. Only escalate if chain partner exists.

Calibration rules:

  • A configuration observation is not a finding. "ATS is disabled" is LOW confidence until you prove sensitive data flows over HTTP.
  • "Exported component" is not a finding. "Exported component that dumps the user database without authentication" is HIGH.
  • "Hardcoded API key" is MEDIUM until you establish whether the key is live and what access it grants. Establish that from the provider's documented privileges for that key type, from the program's own confirmation, or from a test the program authorized in writing against an in-scope asset. Do not exercise a third-party hosted credential without that permission; an unverified key stays MEDIUM rather than becoming an unauthorized test.

Cost Management

The orchestrator tracks token usage and agent spawns:

ControlDefaultPurpose
budget_per_huntFrom providers.jsonTotal token budget (input + output + thinking)
max_agent_spawnsFrom providers.jsonHard cap on total agents
max_loop_iterations3Maximum re-hunt cycles

Effort control: Opus/Sonnet 4.6 use adaptive thinking (thinking: {type: "adaptive"}). Control reasoning depth per agent via the effort parameter (low/medium/high/max). Lower effort = fewer thinking tokens = lower cost. Each agent's effort level is configured per cost profile in providers.json.

Cost-saving strategies:

  • Use Haiku for scope (structured extraction, not reasoning)
  • Use Sonnet for validation and chaining (structured decision-making)
  • Reserve Opus for hunters and report writing (deep reasoning)
  • Skip binary analysis if string extraction + static tools cover the surface
  • Terminate early if 2+ validated findings are ready

Dead End Documentation

When a vulnerability class yields nothing, document it:

{
  "focus_area": "ios_url_schemes",
  "status": "clean",
  "checks_performed": [
    "Extracted URL schemes from Info.plist: myapp://, myapp-share://",
    "Checked for OAuth callback patterns: none found",
    "Checked for sensitive parameter passing: none found",
    "Verified universal links are properly configured with AASA"
  ],
  "reason": "URL schemes used only for non-sensitive deep linking (content sharing). OAuth flow uses universal links with proper AASA configuration."
}

This prevents the agentic loop from re-checking clean areas and helps the orchestrator make informed re-hunt decisions.


Never-Submit Awareness

Before any finding advances to validation, pre-check against the never-submit list:

  1. Missing certificate pinning alone → CHAIN_REQUIRED at best
  2. android:allowBackup="true" alone → CHAIN_REQUIRED
  3. android:debuggable="true" → note to program, not a finding
  4. Missing root/jailbreak detection → informational
  5. Cleartext traffic configuration alone → CHAIN_REQUIRED
  6. Missing ASLR/PIE on modern OS → informational
  7. Self-signed certificates in bundle → often intentional for pinning
  8. Analytics SDK data collection → privacy, not security
  9. Clipboard access on iOS 16+ → system notification mitigates
  10. Missing obfuscation → informational

Hunters should not waste time building full findings for never-submit items unless they already see a chain partner.


Quality Signals

A good hunt produces:

  • 2–5 validated findings per target (quality over quantity)
  • Clear dead-end documentation for covered areas
  • At least one chain attempt if standalone findings are weak
  • Honest confidence ratings (a clean result is better than inflated findings)
  • Reports that pass validator gates on the first attempt

A bad hunt produces:

  • 20+ raw findings that all get killed in validation
  • No dead-end documentation (forcing redundant re-hunts)
  • Inflated severity estimates that get downgraded
  • Generic findings without concrete evidence

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.