Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Use this skill when: planning or executing authorized external recon (red team, bug bounty, ASM); mapping an org's attack surface; investigating a person/entity/threat-actor; producing client deliverables.
Do NOT use this skill when: the user needs active exploitation, post-exploitation, or malware dev; blue-team/detection content; or the target's authorization is unclear — surface the scope question first.
1. Authorization & Legal Posture
Intended for assets the operator owns or has written authorization to assess.
Soft scope check — when authorization isn't established, ask once:
"Quick scope check: is this a target you own or have written authorization to assess? I want to make sure we stay on the right side of the engagement boundary."
Once asserted, don't re-ask. If the engagement type is stated ("pentest of acme.com under contract"), proceed.
Always-on guardrails:
Never weaken auth, rate limits, or safety controls on the target side.
No destructive probes (SYN scans at line-rate, masscan, fuzzing) outside explicit --aggressive mode.
Never paste real PII, credentials, session tokens, or API keys into cloud-hosted LLMs.
Never act against assets outside documented scope, even "obviously related" ones.
2. Confidence Levels
Every assertion carries a confidence level.
Level
Meaning
TENTATIVE
Plausible from indirect evidence; unverified. Snippet-only dork match, email pattern inferred from name, single passive-source subdomain.
Tenant GUID extracted AND domain ties back via MX/autodiscover/SP record
Default reporting posture: never claim CONFIRMED without explicit corroboration. When in doubt, downgrade.
3. Output Format
Each finding uses this schema (drops cleanly into asset-management tools):
Finding:
id: <stable hash or UUID>
module: <technique that discovered it>
asset_key: <typed key, e.g. sub:api.example.com>
category: <e.g. SECRET_LEAK, OPEN_GRAPHQL_API, SSO_EXPOSURE>
severity: <info|low|medium|high|critical>
confidence: <tentative|firm|confirmed>
title: <one-line summary>
description: <2-5 sentences>
evidence:
url: <where found>
timestamp: <UTC ISO8601>
sha256: <hash of any downloaded artifact>
raw: <truncated to 2 KiB>
references: [<CVE-ID, advisory URL, vendor doc>]
remediation: <action the asset owner can take>
Always use UTC timestamps.
4. Source Hygiene & Citations
For every artifact: URL + UTC timestamp + SHA-256 + tool version + run_id.
Hash all downloads with SHA-256. Screenshot in PNG.
Raw HTTP captures capped at 2 KiB body. JSONL logs, one line per event.
Separate evidence read-only from working copies; never edit captured artifacts.
Prefer durable references (CVE, ATT&CK technique ID, RFC). If ephemeral, archive first (archive.today, Wayback SavePageNow).
5. Do NOT
Do NOT paste creds, session tokens, real PII, or unique pivots into cloud LLMs. Use local models for sensitive analysis.
Do NOT assume vendor labels are ground truth (TRM, Chainalysis, Arkham can disagree).
Do NOT assert ownership from a single signal (favicon hash, shared NS, shared CT issuer — each is a hypothesis).
Do NOT run fuzzing, SYN scans, masscan, or nuclei fuzzing/* outside explicit --aggressive mode.
Do NOT use a credential validator for anything except read-only verification.
Do NOT mirror-image the threat actor. Separate capability from intent and sponsorship.
Do NOT escalate when you hit active defenses — back off and document (§6.4).
6. OpSec
6.1 Sock Puppets
Build posting history, age the account, use a separate browser profile. Persona generation: Fake Name Generator, This Person Does Not Exist. Browser isolation: Firefox Multi-Account Containers. Disposable numbers for SMS verification. Audit every extension before install. Maintain chain-of-custody: timestamp every action, hash every artifact.
6.2 Detectability Tagging
Tag every operation so you can reason about the trail you leave.
Tag
Examples
Low
Passive Shodan InternetDB; crt.sh; Wayback CDX; SecurityTrails PDNS; Hunter.io; HTTP HEAD on public buckets; getuserrealm.srf; OIDC metadata fetch.
Active port scans (naabu/masscan/nmap); Nuclei full runs against production; subdomain brute-force at scale; SMTP RCPT TO enum; web fuzzing.
Defaults: passive by default. Active probes only when (a) explicitly authorized, (b) within agreed windows, (c) operator aware of log volume.
6.3 Validator Discipline
When you find a credential in the wild, confirm liveness with read-only validators only (/me, auth.test, sts:GetCallerIdentity). Never create, modify, delete, or send. Record checked_at UTC + truncated response + scope/account-ID. Concrete validator endpoints for 9 providers live in offensive-osint §23.
6.4 Detection-Aware Probing
Signs you've been detected (escalating severity): 429 / Retry-After; captcha interstitials; WAF block page; status-code drift (200→403 from your IP only); banner change; NXDOMAIN rollback; honeypot bait (credentials that don't validate); direct contact.
Back-off ladder:
Halve concurrency; add 2–10s jitter.
Stop hitting the triggering path; pivot to a different module.
New User-Agent / TLS fingerprint.
Rotate egress IP (residential proxy, different cloud region).
Pause 1–24 hours.
If WAF block / status drift / direct contact: stop and consult the engagement lead.
7. External Red-Team Recon Pipeline
Six sequential operational stages, then reporting as a post-stage; modules within a stage can run concurrently.
Stage
What you do
1 — Seed Discovery
WHOIS, ASN enum (HE BGP Toolkit, RIPEstat), DNS records (A/AAAA/MX/TXT/NS/SOA/CAA), CT history (crt.sh, Censys).
2 — Asset Expansion
Subdomain enum (passive first → permutations → brute); cloud bucket permutation; typosquat generation; Wayback CDX; mobile app discovery; DNS walking; LinkedIn employee enum; org/subsidiary footprint (reverse-WHOIS, corporate registries).
One bounded recursion round over hostnames/assets surfaced by Stages 2–4 (re-mutate, re-probe — not an open-ended crawl); cloud-footprint follow-up (e.g. account-ID extraction from a leaked cloud access key); supply-chain claimability check (internal/scoped package names confirmed unclaimed on the public registry).
6 — Active Validation
Hard-gated, operator-armed, engagement-authorized proof tier — default OFF. Live credential submission (default-cred + breach-credential replay) against owned panels; active cloud/container/API/SSO confirmation; host-header injection / blind-SSRF; HTTP request-smuggling differential; injection-payload confirmation (e.g. XSS); JWT-forge auth-bypass differential. Every module here needs explicit arming plus a per-target scope-confirmation step on top of engagement authorization (§1).
Profiles: 1-hour rapid (Stages 1–2 passive + breach + exec summary) · 4-hour focused (adds email harvest, SSO fingerprinting, typosquats) · 1-day standard (full Stages 1–5 in priority order) · 1-week deep (all of standard + JS deep, mobile, cloud-native, vendor product, package registry; Stage 6 only under explicit written engagement authorization) · ongoing weekly diff (re-run Stages 1–3, diff against baseline).
Abort conditions: scope mismatch after Stage 1; near-zero attack surface after Stage 2 — meaning the union of every connector in a category came up empty, not just one source (§7.3); WAF/detection signs hit during any stage (§6.4).
7.3 Connector Resilience
No single data source should gate a whole category. Query interchangeable backends per category as a union, not a single call — CT logs (crt.sh, Censys, certspotter), passive DNS (multiple providers), host-intel (Shodan, Censys, Netlas), reverse-WHOIS (multiple vendors) — with per-source health tracking and short-TTL result caching underneath the union.
A rate-limited or dead source degrades that category's yield; it must never silently zero it out.
Log the failure as a per-source health event, not as "category empty." The Stage-2 abort decision (§7.2) reads the union result, never one source's status.
Cache each source's results with a TTL short enough to stay fresh but long enough to survive that source's rate-limit window — a re-run mid-engagement shouldn't re-spend a quota you already burned.
This generalizes the crt.sh-style fallback discipline to every category, not just certificates: one dead API is a coverage-degradation event to note and route around, not a hole in the map.
7.4 Stage Number ≠ Safe-by-Default
A module's stage number tells you when it runs, not whether it's passive. Always check the module's own gating flag before treating a stage placement as a detectability tier.
Technique
Stage
What actually gates it
IDOR/BOLA differential confirm
3 (enrichment)
its own --active/--deep flag — not the stage boundary
Active user-enumeration oracle (IdP)
3 (enrichment)
its own --deep flag
Bounded 401/403 bypass / param mining
4 (exposure)
rate-limited by default; ships un-gated at this stage
Credential replay / default-cred submission
6 (active validation)
explicit arming + scope confirmation
JWT-forge auth-bypass differential
6 (active validation)
explicit arming + a second confirmation flag
Stage 6 is the one place the pipeline is explicitly all-intrusive-by-design (default OFF, operator-armed) — but stages 1–5 are only mostly passive-to-low-intrusion, and "mostly" hides real exceptions like the two Stage-3 rows above. Detectability is tagged per-technique (§6.2), never inferred from the pipeline stage a technique happens to sit at.
8. Asset Graph Discipline
Every discovery is a typed asset in a graph, not a free-floating string.
Discipline: create the asset first, then attach the finding. Dedup by key. sources[] must list every source. Confidence is per-source, then aggregated.
Repo priority: recently pushed (last 30 days) > stale; public with target name in description > code-only; mentions prod/internal/secret in name → HIGH priority despite being public.
9. Findings Rubric & Severity Mapping
Severity Anchors
Severity
Anchor
CRITICAL
Pre-auth code execution; confirmed valid credentials; listable production data; fundamental trust violations. Examples: .env exposed, listable S3 bucket with PII, live-validated AWS admin key, open Kubernetes API with anon-auth, ≥10 employees in breach corpus + tenant identified.
HIGH
Significant exposure with clear escalation path; high-value info disclosure. Examples: public secret in GitHub repo, subdomain takeover possible, reflected CORS with credentials, exposed Jenkins/phpMyAdmin admin UI, open GraphQL introspection on prod, DMARC p=none.
MEDIUM
Info disclosure, hardening gaps, brute-force exposure. Examples: missing HSTS/CSP, Apache /server-status, internal IP/hostname in JS, schema leakage in error pages, android:allowBackup=true, wildcard CORS on user-data API, Slack webhook leaked.
LOW
Cosmetic or marginal gaps. Examples: missing X-Frame-Options, .DS_Store exposed, Stripe test key, cert pinning missing, outdated WordPress (no known active exploit).
INFO
Worth recording; no immediate action. Examples: robots.txt reveals paths, private bucket locked down, DNSSEC not enabled.
Severity Escalation Rules
HSTS missing on auth/login/SSO/admin path → MED → HIGH.
Wildcard CORS + credentials header → MED → HIGH.
Endpoint interest score ≥70 (companion skill §20) → at least HIGH.
Domain breach ≥10 employees → CRITICAL regardless of stale-data caveats.
Vendor product version matches CISA KEV → CRITICAL.
10. Pivot Modes & Scale Tactics
Aspect
Investigative Mode
Offensive Recon Mode
Probing rate
Slow, single-threaded, blends with traffic
Bursts, parallel, rate-limited per provider
OpSec posture
Sock-puppet only; never reveal investigator
Engagement persona; team may notify SOC
Evidence handling
Court-grade chain of custody
Engagement-grade; same hash/timestamp discipline
Reporting format
Narrative + sourced timeline
Per-asset findings + remediation + reproduction
Scale tactics:
Small (<100): Individual-account focus. One exec/CFO compromise often hands you the keys. Deep on every email + every identity-fabric finding. Check founders' personal GitHub orgs.
Medium (100–1K): Balanced enumeration. Full pipeline at standard depth. LinkedIn priority by role. Check both app stores.
Large (1K–10K): Breadth-first; automation for asset discovery; manual triage on findings only.
Very large / conglomerate: Scope pruning is the most important step. Brand-pivot map first. Breach corpus and systemic posture findings (DMARC gaps, SSO_EXPOSURE breadth) dominate over individual accounts.
11. Implementation: Companion Skill Pointers
The following modules have full implementation detail — probe paths, wordlists, curl one-liners, regexes, and scoring rubrics — in offensive-osint. This skill defines what to do; that skill defines how to do it.
Phishing Infrastructure (offensive-osint §16.14 for email security) — typosquat shortlists via dnstwist; subdomain takeover for trusted-domain phishing; email spoof feasibility matrix (SPF × DMARC); pretext development from OSINT (job titles, recent events, vendor relationships, GitHub commits).
12. Breach × Identity Correlation
Highest-ROI single technique for external red teams. Run on every engagement.
Source
Tier
Notes
Hudson Rock Cavalier
FREE
Infostealer-log corpus; very high signal for corp SSO creds.
Have I Been Pwned
Free + paid
Domain-wide existence + Pwned Passwords (k-anonymity).
DeHashed
Paid
Per-record searchable API.
IntelX
Free + paid
Aggregator; phonebook search.
Domain-level severity: ≥10 employees compromised → CRITICAL; 1–9 → HIGH; ≥1 end-user → MEDIUM; domain seen with 0 named accounts → INFO.
SSO_EXPOSURE: after Stage 3 identity-fabric mapping AND breach lookups, intersect discovered IdP tenant domain with breach corpus. Non-empty intersection → SSO_EXPOSURE finding, severity CRITICAL. Evidence: tenant ID + product + employee count + per-account source.
Stealer log discipline: encrypt at rest; SHA-256 every artifact; never paste plaintext passwords into cloud LLMs; maintain chain of custody; redact passwords in client reports by default (offer encrypted credential bundle separately).
12.1 Per-Person Identity Dossier
Domain-level breach counts (above) tell you the org is exposed. The identity dossier tells you who — join every discovered PERSON (name, role, seniority, target score from person-enum/LinkedIn dorking) to any leaked CREDENTIAL matched by email, producing one ranked card per human. This is the step that turns a breach headcount into a spear-phish / credential-stuffing target list.
The join key is the email address, not the name. A PERSON asset is name-keyed and usually carries no email on its own — resolve name→email via the org's confirmed email-format pattern (Hunter-derived or independently observed) before joining. Never join off a guessed pattern from a single sample.
Field
Source
Role / seniority / target score
person-enum / LinkedIn dorking
Email — known (directly observed) or pattern (name run through the org's confirmed format)
Risk tiering: a breached high-value mailbox (exec/IT/admin, target score ≥80) with ≥1 confirmed credential is the worst case.
Target score
Credential(s) leaked
Risk
≥80 (exec/IT/admin)
≥1
CRITICAL
any
≥1
HIGH
≥80
0
MEDIUM
<80
0
LOW
Orphan credentials: a leaked email that ties to no discovered person is still a lead, not a dead end — surface it separately, sorted by breach count. It just hasn't been named yet; a further person-enum pass may resolve it.
Confidence: the join is only as strong as the email resolution. known email carries the credential's own confidence (§2); a pattern-resolved email stays TENTATIVE for attribution purposes until the address itself is independently observed (e.g. it appears verbatim in the breach record, or Hunter.io returns it directly) — never report a pattern-resolved dossier as CONFIRMED off the pattern alone.
Run this pass any time Stage 3 person-enum/email-OSINT output and a breach lookup (domain-level or replay) both exist for the engagement — it is a join over data already collected, not new collection, so there's no reason to skip it. Use the output to decide who gets the phishing pretext first (§13) and who's the highest-value candidate for the read-only credential check (§6.3).
13. Specialty OSINT Domains
Cryptocurrency — track flows with Cielo, TRM, Arkham, MetaSleuth. L2/rollup: start at L1 bridge events; use L2 explorers for in-rollup activity. Caution: bridges mint/burn (avoid 1:1 flow assumptions); MEV paths create false direct trails.
Image / Video / Chronolocation — reverse image search (Google Lens, Yandex, TinEye); EXIF via ExifTool; forensics via Forensically/FotoForensics; geolocation via foreground+background landmark analysis, Street View, Overpass Turbo, PeakVisor. Shadow analysis: SunCalc, ShadeMap. Satellite: Google Earth Pro historical, Sentinel Hub.
Threat Actor Investigation — scoping: actor hypothesis from CERT/vendor reports → IOC harvest → infra mapping via CT log pivots, shared hosting, NS reuse, HTML fingerprints → artifact profiling (PDB paths, Rich headers, SSDEEP/YARA) → social pivots (handles, code snippets, job posts). Attribution discipline: rule of three; separate capability from intent; prefer durable pivots (code-signing certs, build path idioms) over ephemeral (resolving IPs). Russia pivots: EGRUL, Rusprofile, hh.ru, VKontakte. China pivots: gsxt.gov.cn, Tianyancha, ICP filings, Weibo, Zhihu.
People & Social Media — username enumeration: WhatsMyName, Sherlock, Maigret. Face search: PimEyes, Exposing.ai. Social graph: Maltego, SocialBlade. Bluesky: DID resolution via bsky.social/xrpc/, firehose via Firesky. Mastodon: WebFinger discovery; FediSearch cross-instance.
14. Anti-Patterns & Common Failure Modes
Single-source attribution. Rule of three.
Trusting vendor labels as ground truth. Labels are hypotheses.
Favicon-hash = ownership. Shared infra, shared CMS, shared CDN all produce matches.
Snippet-only dork as CONFIRMED. TENTATIVE until visited.
Pasting real PII / creds into cloud LLMs. Local models only.
Mirror-imaging the threat actor. They don't think like you.
Attribution by IP geolocation. VPNs and residential proxies exist.
Ignoring CT-log lag. Absence ≠ doesn't exist; lag can be minutes to hours.
Counting Wayback as "the site at time T." Best-effort; many requests fail.
Letting the asset graph carry untyped strings. Every discovery is an asset.
Skipping the scope check. Ask once when in doubt.
Forgetting UTC. Local time creates correlation bugs.
Continuing to probe after a WAF block. Back off (§6.4).
Skipping confidence-upgrade documentation. TENTATIVE needs a path to CONFIRMED.
Treating exec-summary as an afterthought. Plan deliverables at engagement start.
Title: [Severity] [Component] Brief description
Summary: 2-3 sentences — what and why it matters.
Steps to Reproduce: numbered, copy-pasteable, URL + payload + expected vs actual.
Proof of Concept: screenshot or sanitized HTTP request/response.
Impact: what data/users/functions are at risk.
Severity: CVSS v3 vector + score + 1-sentence justification.
Remediation: concrete, actionable recommendation.
Unprogrammed CVD: check security.txt → security@<target> → WHOIS abuse contact → CERT/CC. Standard 90-day window before public release. Never: include others' PII, post publicly before window expires, or escalate via social media first.
16. Client Deliverable Templates
Executive summary structure: engagement metadata → top 3–5 findings (title + business impact + remediation effort) → postural observations (email security, identity fabric, cloud surface, mobile) → aggregate metrics (assets, findings by severity, live creds confirmed) → recommended next steps with timeline.
"Run full subdomain enum on chase.com." → §1 (scope check; should NOT run).
"crt.sh is 429ing mid-engagement — is the subdomain list complete?" → §7.3.
"A module runs at Stage 3. Safe to assume it's passive?" → §7.4.
"We have breach hits and a LinkedIn employee list — who do we prioritize?" → §12.1.
18. Changelog
v2.3 (2026-08-06) — §7 upgraded from the 5-stage model to the battle-tested 6-stage model (Seed → Asset Expansion → Enrichment → Exposure Analysis → Convergence [bounded recursion + cloud-footprint + supply-chain] → Active Validation [hard-gated, operator-armed, default OFF]), with Reporting reframed as a post-stage; added §7.3 Connector Resilience (union-query interchangeable backends per data category, per-source health tracking, so one dead/rate-limited source never zeroes a category) and §7.4 Stage Number ≠ Safe-by-Default (a module's stage doesn't tell you if it's passive — always check its own gating flag; detectability is per-technique). Added §12.1 Per-Person Identity Dossier — joins discovered PERSON role/seniority to leaked CREDENTIAL by email into a ranked spear-phish/credential-stuffing target list, with risk tiering and orphan-credential handling. Added 3 self-test prompts (§17.13–15).
v2.2 (2026-04-29) — refactor: trimmed from 1,694 to ~480 lines. Compressed implementation-detail sections (§11–§15, §27–§31 original) to pointers to offensive-osint. Retained full framework core: confidence levels, pipeline, asset graph, severity rubric, OpSec, breach correlation, anti-patterns, deliverable templates. Removed duplicate content; combined specialty domains into single §13; merged §23–§25 into §13; collapsed §27–§29 into §11 pointer block.
v2.1 (2026-04-27) — comprehensive expansion based on 32-prompt smoke-test gap analysis. PASS rate: 31/32.
v2.0 (2026-04-27) — major rewrite for external red-team posture.
v1.x — original framework based on SnailSploit/offensive-checklist.
