PerimeterX / HUMAN Security — SDK Reverse Engineering

Pure-algorithm reconstruction of PerimeterX's _px3 / _px2 cookie chain — no browser, no automation framework.

中文 · Quick Start · How it works · Docs · AI Skill · License

Given a target URL + AppID, this project rebuilds the PerimeterX collector POST chain entirely in pure math (Node.js) and emits a valid _pxN cookie. Every constant, field position, and field type is extracted from live captures, never from memory or stale docs. Research / education only.

⚠️ Research, education, and personal security-audit use only. Not a scraping tool — ships no proxy pools, schedulers, IP rotation, or solving integrations. See Disclaimer before use.

What it does

• Reconstructs the silent collector flow — EV1/EV2/EV3 sensor payloads → collector POST chain → _pxN, no headless browser.
• 9 reusable crypto/encoding primitives (revers/) — payload XOR/base64/interleave, PC (HMAC-MD5), OB decode, SID stego, UUID v1, djb2, memory, anti-tamper, /ns.
• Template methodology — capture 6 real batches, lock STATIC fields, reverse DYNAMIC fields from the SDK. ~90% of code ports across sites; only constants + a handful of b64 keys differ.
• Three deployment tiers — the same SDK is configured differently per customer; the project documents lenient → strict → strict+ and what each extra server-side check requires.

Validated sites

Site	Tier	Cookie	Chain	Result
ifood.com.br	lenient	_px3 (ttl 330)	2-POST	10/10
grubhub.com	lenient	_px2 (ttl 500)	2-POST	10/10
walmart.com	lenient	_px3 (ttl 330)	2-POST	10/10 ²
totalwine.com	strict	_px2 (ttl 330)	3-POST + EV3	10/10
academy.com	strict+	_px3 (ttl 330)	3-POST + TLS/IP-bound trust	10/10 ¹

All constants come straight from real POST-body captures (6 auditable batches per public site). ¹ academy is a strict+ case (trust bound to the mint's TLS fingerprint, /ns token, real-Chrome template, and exit-IP reputation); its generator package is kept private. The tier methodology is documented in the AI skill. ² walmart is PX-layer only — the site's primary gate is Akamai Bot Manager, with PerimeterX as a secondary layer. The _px3 is accepted by the PX collector (10/10) but does not pass Walmart business APIs (Akamai _abck/bm_sv is out of this skill's scope). Included as a clean lenient + cross-event-consistency reference.

Quick start

git clone https://github.com/warterbili/PerimeterX_RE.git
cd PerimeterX_RE && npm install

# iFood — generate _px3 (lenient tier, runs as-is)
node stample/ifood/px_cookie/ifood_px3.js
# → ✅ _px3=eyJ1IjoiYWJj…  ttl=330   ev1=14  ev2=204

# Grubhub — generate _px2
node stample/grub/px_cookie/grubhub_px2.js

# Walmart — generate _px3 (lenient; PX-layer only, needs a US egress IP)
node stample/walmart/px_cookie/walmart_px3.js
# → ✅ _px3=a2426f96…  ttl=330  ev1=12 ev2=207
# ⚠️ PX-layer only: Walmart's business APIs are gated by Akamai (out of scope).
# PX-success differential proof: node stample/walmart/script/px_success_proof.js

# Total Wine — generate _px2 (strict tier, 3-POST; needs a US residential proxy)
HTTPS_PROXY=http://user:pass@host:port node stample/totalwine/px_cookie/totalwine_px2.js
# → _px2=…  ttl=330  ev1=13 ev2=199 ev3=11  seq2=200

# End-to-end business call (proxy required) — cookie → real gated HTML
node stample/totalwine/px_cookie/business_api_demo.js

Each stample/<site>/px_cookie/ has a smoke_test.js (offline self-check) and business_api_demo.js (end-to-end). The press-challenge path is a userscript — see bundle/.

How it works

PerimeterX runs two defenses. 99% of traffic is the silent collector: the SDK serializes sensor events (EV1/EV2/EV3), encrypts them (XOR → base64 → interleave), signs a PC (HMAC-MD5 digit extraction), POSTs to the collector, and gets back an OB-encoded _pxN. The <1% press-challenge path adds a PoW + WASM bundle — handled separately in bundle/.

A generator is a real capture template with only the dynamic fields overwritten: server state (parseInt'd), HMAC/MD5 fields (inputs reversed per-site from the SDK), counters, timestamps, the /ns token, and the anti-tamper slot. Strict tiers add a mandatory EV3 cookie-confirmation beacon, counter sync, and hid; strict+ additionally binds trust to the mint transport (real Chrome TLS) and exit-IP reputation. Full methodology: main/ and the AI skill.

Project structure

revers/        9 algorithm modules (require()-able)         skill/      AI reverse-engineering skill
stample/       per-site generators + 6 capture batches        ├─ AI_re/  PX core skill (playbooks + tools)
  ├─ ifood/    lenient · _px3                                  └─ cdp/    real-Chrome CDP capture skill
  ├─ grub/     lenient · _px2                                main/       technical docs (ZH + EN)
  ├─ walmart/  lenient · _px3 · PX-layer only (Akamai-gated) bug_report/ gotchas / failure-mode records
  ├─ totalwine/ strict · _px2 · 3-POST                       research/   field-entropy + SDK-drift studies
  └─ academy/  strict+ · _px3 · TLS/IP-bound trust
bundle/        press-challenge path (PoW + WASM userscript)
node_bridge/   JSDOM "Plan B" oracle (run the real SDK)

Documentation

Doc	What
main/ZH · main/EN	Full technical reference — algorithms, protocol, EV field tables, cross-version grep methodology
skill/AI_re/	The reverse-engineering skill: 7 references, 11 playbooks, 17 CLI tools, 23 gotchas
skill/AI_re/references/validated-sites.md	Per-site constants + b64-key maps + tier checklist (read before a new site)
skill/AI_re/references/deployment-tiers.md	Lenient / strict / strict+ comparison
bug_report/	Documented failure modes / pitfalls

Limitations

• Tied to a locked SDK version per site; when PX ships a new SDK, re-capture and re-derive (the skill has an upgrade playbook).
• Strict+ sites need a real-Chrome-derived template + Chrome-TLS transport + a clean residential exit IP per cookie — a byte-correct cookie minted over node TLS or a degraded IP is still challenged.
• The press-challenge (PoW/WASM) path is a browser userscript, not pure-math.

License

Dual-track (full text in LICENSE):

Asset	Scope	License
Code	revers/ · stample/*/px_cookie/ · bundle/script/ · node_bridge/ · skill/*/scripts/	AGPL-3.0 — commercial/SaaS use must open-source back
Docs	main/ · bug_report/ · research/ · all README / SKILL.md	CC BY-NC-SA 4.0 — non-commercial + attribution + share-alike

Ethics & Disclaimer

Research / education / personal-audit only; all capture samples were collected via the researcher's own accounts. Must not be used for unauthorized scraping, carding / coupon fraud, scalping, credential stuffing, DoS, or anything violating a target's ToS or local law (US CFAA / EU GDPR / China CSL & PIPL, etc.). The author releases this as an academic artifact and assumes no liability for any resulting account ban, IP blocklisting, litigation, or loss; legal compliance is the user's responsibility. Redistributions must keep the full disclaimer in LICENSE. iFood / Grubhub / Total Wine are protocol-research objects only — the project neither encourages nor instructs any malicious action against them.

Responsible disclosure — all findings come from protocol- and algorithm-layer reverse engineering: no internal source, private APIs, or unauthorized access; no backend-vuln exploitation; no TLS/cert/signature bypass; no theft of third-party keys or data. Conforms to reverse-engineering fair use and security-research exceptions (e.g. US DMCA §1201(j)). Compliance contact: warterbili via GitHub Issues — the author reserves the right to adjust disclosure scope within 90 days of a formal request.

Citation

@misc{perimeter_v2_2026,
  author       = {warterbili},
  title        = {{PerimeterX (HUMAN Security) SDK Reverse Engineering}},
  year         = {2026}, version = {2.0},
  howpublished = {GitHub Repository},
  url          = {https://github.com/warterbili/PerimeterX_RE}
}

_{v2.0 · research artifact · back to top}