Browser Browser

Сканирование

Новый скан Мониторы

Инструменты

My IP DNS Lookup WHOIS SSL Certificate Ping HTTP Headers Domain Check IP Calculator IDN Converter Reverse DNS Schema Generator TAS-IX Трассировка

Сервисы

Массовый скан Хостинг Отчёт CSEC Дефейс

Разведка

Армия AI-агентов Конкуренты Граф CVE Страхование Репутация

SEO

AI Blog GitHub SEO Site Audit Site Compare Traffic Analytics

Ещё

Все функции Документация Цены
Начать бесплатно
7-STAGE · PARALLEL · ADVERSARIAL

Specialised hunters,
not one big agent.

Eight parallel hunters, each specialised in a single vuln class.

Run the harness → See the 7 stages
Public-repo & raw-source input · No exploitation · ~$0.30-1.50 per run
Why "one big agent" fails on real codebases
ANTI-PATTERN
Single end-to-end agent
One agent pointed at a 100k-line repo covers roughly 0.1% of the attack surface usefully before its context window fills up. Vulnerability research is parallel and narrow — sweeping the whole codebase serially is the wrong architecture.
RIGHT SHAPE
Many narrow hunters in adversarial dispute
Each hunter sees a small slice through one vuln-class lens. A separate adversarial validator with different prompting tries to disprove every claim. Cloudflare's note: two agents in conscious dispute cut more noise than telling one agent to "be careful."

The seven stages

1/7
RECON
Map trust boundaries, entry points, sensitive sinks
2/7
HUNT ×8
Parallel hunters: SQLi · CMDi · SSRF · deser · XSS · authn · crypto · info-leak
3/7
VALIDATE
Adversarial agent tries to disprove every finding
4/7
GAP-FILL
Re-hunt sinks flagged but uncovered
5/7
DEDUPE
Collapse findings sharing one root cause
6/7
TRACE
Taint-check: is attacker input actually reaching the sink?
7/7
REPORT
Structured, ranked, reachability-tagged output

Sample run — github.com/example/webapp

browser.uz/mythos · run #142 · GPT-4o-mini
[01] RECON → complete · 7 entry, 12 sinks [02] HUNT → firing 8 parallel hunters sql_injection → 4 findings command_injection → 2 findings ssrf → 1 finding deserialization → 0 findings xss → 3 findings authn_authz → 2 findings crypto_memory → 1 finding info_disclosure → 2 findings [03] VALIDATE → adversarial · 11 survived · 4 refuted [04] GAP-FILL → 1 extra finding on un-covered sink [05] DEDUPE → 12 → 9 unique root-causes [06] TRACE → 6 reachable from entry · 3 unreached [07] REPORT → sorted & emitted CRIT SQL injection in /api/search?q= · [reachable] CRIT SSRF in /admin/webhook URL field · [reachable] HIGH Reflected XSS in /comments rendering · [reachable] HIGH JWT signature bypass in /auth/verify · [reachable] HIGH Path traversal in /static/file?n= · [reachable] MED Weak crypto (MD5 for password reset) · [reachable] MED Stack-trace exposure on error · [unreached] MED Hardcoded secret in legacy module · [unreached] LOW Verbose error logs (info disclosure) · [unreached]

vs. "wrap GPT around Semgrep"

Aspect Browser Mythos GPT + Semgrep wrapper
Parallel narrow hunters?✓ 8 specialised✗ one generic pass
Adversarial validation?✓ disprove-prompt✗ no
Reachability trace?✓ taint-check✗ no
Root-cause dedupe?✓ collapsed✗ raw dump
Gap-fill re-queue?✓ second pass on uncovered sinks✗ no
Live stage progress?✓ 7 stage indicators— spinner
Cost per run~$0.30-1.50$5+ from token waste
⚠ Honest take — Mythos preview, not Mythos prod

Honest scope: same architectural pattern at a sustainable cost, not a 1:1 clone.

Run the harness on your repo.

Paste a public GitHub URL or your source code. 7 stages later, a structured findings report — deduplicated, validated, reachability-tagged.

Start a run →