ADVERSAI
Start free
v3.1 shipped · 458 tests greenautonomous · scope-boundUAE federal · licensed offensive

We break in
before they do.

Now with deep content extraction — 14 engines that pull actual evidence, not just warnings.

Five reasoning engines — OSINT, Network, Cortex, Phantom, LLM Red-Team — plus eight Tier-1 deep-extraction engines, driven by a ReAct loop, scope-validated on every call, and audited against a three-pillar UltraReview before a single claim ships.

Start a scanSee a live scan
tests green
458/458
attack engines
14
AI agents
13
secret classes
20+
the problem

Attackers run 24/7.
Defenders run once a quarter.

  • 01

    Manual pentests ship too late

    A human engagement runs twice a year and covers a fraction of your attack surface. Attackers don't scope.

  • 02

    Scanners are loud and shallow

    Commodity scanners read headers and call it a day. They miss CORS reflection, DKIM drift, LLM injection, ASM deltas.

  • 03

    LLM apps are a new attack surface

    Jailbreaks, prompt injection, tool abuse, retrieval leaks. Most tools can't even see them, let alone test them.

  • 04

    Guardrails are mostly paper

    CLI caps that don't cap. Scope flags that never reach the engine. Paper guardrails convert into production bills.

the stack · core

Five reasoning engines. One ReAct loop.

Every call routes through the same scope validator. Every finding routes through the same ReAct reasoning trace. Every trace is replayable on a network-blocked provider, deterministically.

the stack · v3.1 deep extraction

Eight new engines. Evidence, not warnings.

Tier 1 runs pre-ReAct on every scan. Each engine pulls actual extracted content from the target — bundle contents, buckets, schemas, historical infra — and writes the evidence to disk next to the finding.

  • tier 1 · 01

    JS Bundle Miner

    Scans webpack/vite bundles for 20+ secret classes (AWS, Stripe, GitHub, Supabase, Firebase, JWTs).

    hoppa.global — 715KB bundle beautified, regex-scanned, zero leaks
  • tier 1 · 02

    Source Map Reconstructor

    Rebuilds full source code when .js.map files ship to prod — 14 URL patterns probed per target.

  • tier 1 · 03

    GraphQL Introspector

    Dumps complete schema, detects exposed playgrounds (GraphiQL / Apollo), writes SDL + JSON.

  • tier 1 · 04

    .well-known Enumerator

    Harvests robots.txt, OIDC config, security.txt, sitemap — parses Disallow paths for sensitive surface.

  • tier 1 · 05

    Subdomain Deep Scanner

    Certificate transparency + DNS brute + amass passive → full historical subdomain inventory, live-host classified.

    hoppa.global — staging. + test. surfaced as HIGH
  • tier 1 · 06

    Bucket Enumerator

    Finds exposed S3 / GCS / Azure buckets by brand pattern, parses XML listings, captures up to 100 keys.

  • tier 1 · 07

    Full Port Scanner

    nmap top-100 with service detection. Risky ports (MSSQL / Postgres / Redis / Mongo / K8s API) → CRITICAL.

  • tier 1 · 08

    DNS History Miner

    SecurityTrails + VirusTotal + DNSDumpster → every IP the target ever used → finds forgotten legacy infra.

Dogfooded on hoppa.global — 24s wall-clock, 21 findings, 2 HIGH (staging & test subdomains surfaced). 458 tests green, 75% engine coverage. Version 3.1.0.

jurisdiction

Offensive security, licensed in-region.

AdversAI operates under a UAE federal offensive-security authorization. For GCC clients — federal, financial, critical-infra — that means engagements stay jurisdictionally clean: license in-region, execution in-region, evidence in-region.

  • 01
    UAE-licensed offensive security

    Operating under a live UAE federal information-security license. Engagement letters signed in-region, offensive work authorized by national framework — not a re-sold US license or a self-declared statement.

  • 02
    GCC-approved. Federal-scope authorized.

    Scope approved for Gulf Cooperation Council entities, including federal agencies and regulated sectors (finance, telco, critical infra). Cleared to test production systems, not just staging sandboxes.

  • 03
    Data sovereignty — findings stay in-region

    All scan artefacts, extracted secrets, replay caches, and reasoning traces live on Tailscale-private infra inside the UAE. No third-party SaaS ingestion. No cross-border transfer of your evidence.

license number available on request under NDA

the ultrareview

Every claim audited.
Every guardrail honest.

Three parallel adversarial agents audit every major commit — Security, Correctness, Quality — with grep-verified findings and file/line citations. When we find a paper guardrail in our own code, we publish the fix before the changelog.

passed · 2026-04-17
why it matters

Security

Hunts for injection, auth bypass, and dead-plumbing risk in our own code before it reaches yours.

No eval, exec, unsafe YAML, or shell=True anywhere in the codebase.
CSP with strict-dynamic + 144-bit CSPRNG nonce per request; unsafe-eval dropped entirely.
_block_network patches socket/getaddrinfo/gethostbyname — replay mode proven hermetic.
why it matters

"A pentest platform that can be tricked into exfiltrating its own customer data is worthless."

AdversAI is built by operators who've watched automated pentest tools ship paper guardrails — CLI flags that look like caps but never reach the engine, CSP nonces that never get consumed, test suites that monkey-patch the very signature the production code breaks. Every time we catch one of those in our own code, we publish it: the finding, the file, the line, the fix.

— adversai operating doctrine
what you get back

Findings ship as Markdown, JSON, and SARIF.

Reproduction steps, remediation, CVSS, replayable trace. Drop straight into PRs, Jira, or your SIEM.

live demo

Real scans. Real output. No mockup.

These are verbatim traces from production scans, anonymized. Tap an engine to watch it run.

passive recon on acme.example
adversai — osint.trace● hermetic · replay mode
the landscape

What the field actually looks like.

We read the competitors' READMEs, ran their demos, and ported their best ideas (Shannon-style reasoning, PentAGI's autonomy arc). Here's an honest grid. Where a rival is better, we say so.

AdversAI v3.1

ours
Deep content extraction
Yes (8 engines)
Secret extraction from bundles
Yes (20+ classes)
OSINT engine
Yes
Network engine
Yes
Reasoning loop
Yes (ReAct)
Autonomous exploit
Yes (playbook)
LLM red-team
Yes
Scope guardrails
Per-call validator
Black-box (no source)
Yes
Replayable trace
Yes
Licensed / in-region
UAE federal
Providers
4 + Bedrock

Shannon Lite

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
No (white-box)
Network engine
Yes
Reasoning loop
Yes (2-stage)
Autonomous exploit
Yes (PoC req'd)
LLM red-team
No
Scope guardrails
Implicit (source)
Black-box (no source)
No
Replayable trace
No
Licensed / in-region
No
Providers
Claude only

Pentest Swarm AI

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
Yes
Network engine
Yes
Reasoning loop
Yes (ReAct)
Autonomous exploit
Yes
LLM red-team
No
Scope guardrails
Yes
Black-box (no source)
Yes
Replayable trace
No
Licensed / in-region
No
Providers
Claude only

Strix

Deep content extraction
Partial
Secret extraction from bundles
Partial
OSINT engine
Yes
Network engine
Yes
Reasoning loop
Yes (agentic)
Autonomous exploit
Yes
LLM red-team
No
Scope guardrails
Yes
Black-box (no source)
Yes
Replayable trace
No
Licensed / in-region
No
Providers
Multi

Artemis

Deep content extraction
Partial
Secret extraction from bundles
No
OSINT engine
Yes
Network engine
Yes
Reasoning loop
Partial
Autonomous exploit
No
LLM red-team
No
Scope guardrails
Yes (CERT-PL)
Black-box (no source)
Yes
Replayable trace
Partial
Licensed / in-region
PL national
Providers
N/A

METATRON

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
Partial
Network engine
Yes
Reasoning loop
Yes (multi-step)
Autonomous exploit
Partial
LLM red-team
No
Scope guardrails
Yes
Black-box (no source)
Yes
Replayable trace
No
Licensed / in-region
No
Providers
N/A

PentAGI

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
Yes
Network engine
Yes
Reasoning loop
Yes
Autonomous exploit
Yes
LLM red-team
No
Scope guardrails
Yes
Black-box (no source)
Yes
Replayable trace
No
Licensed / in-region
No
Providers
Multi

PyRIT

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
No
Network engine
No
Reasoning loop
No
Autonomous exploit
No
LLM red-team
Yes
Scope guardrails
N/A
Black-box (no source)
Yes
Replayable trace
Partial
Licensed / in-region
No
Providers
Multi

garak

Deep content extraction
No
Secret extraction from bundles
No
OSINT engine
No
Network engine
No
Reasoning loop
No
Autonomous exploit
No
LLM red-team
Yes
Scope guardrails
N/A
Black-box (no source)
Yes
Replayable trace
Partial
Licensed / in-region
No
Providers
Multi

Sources: Shannon Lite (GitHub KeygraphHQ/shannon), Pentest Swarm AI (Armur-Ai/Pentest-Swarm-AI), Strix (usestrix), Artemis (CERT-PL/Artemis), METATRON (agentic-pentest), PentAGI (vxcontrol/pentagi), PyRIT (Microsoft), garak (NVIDIA). Verified 2026-04-18.

Our differentiator: UAE federal license + deep content extraction (20+ secret classes, source-map reconstruction, historical-infra enumeration) + evidence-grade reports (Markdown, JSON, SARIF) with replayable reasoning traces.

unified scoring

One number.
Grounded in five.

Every finding is graded by engine-local severity, then fused into a single composite score weighted by attacker cost, exploitability, and blast radius. The dial on the right tracks the score as you read this section.

  • 0–39Hardened. Minor posture notes only.
  • 40–74Findings exist. Patch path is clear.
  • 75–100Active exposure. Mitigate before it costs you.
0RISK SCORE
pricing

Start free. Grow when you need to.

The Solo tier ships the full kit — you're trusted to run responsibly. Paid tiers add the glue teams actually pay for: shared scope, audit logs, SSO, incident paths.

Tier 2 authenticated testing and Tier 3 full red team engagements require a signed engagement letter — contact us.

  • Solo
    $0self-hosted

    One operator. Full kit.

    • All 5 engines, unrestricted
    • CLI + JSON/Markdown reports
    • Bring your own LLM (Claude, OpenAI, Ollama, LMStudio)
    • Replay cache + non-mocked test harness
    • Community Discord
    Get the code
  • most picked
    Team
    $499per month · 5 seats

    Shared scope, shared reports.

    • Everything in Solo
    • Shared scope registry + audit log
    • Playbook library + ASM watcher
    • HackerOne / Bugcrowd scope sync
    • Priority Slack support, business-hours SLA
    Start trial
  • Enterprise
    Customannual

    SOC-aligned. On your infra.

    • Everything in Team
    • SSO, SCIM, RBAC
    • Private deployment · AWS Bedrock preview
    • Dedicated scope-review engineer
    • 24/7 incident path · named TAM
    Book a walkthrough
  • Research
    Free*reviewed per request

    Academics, CERTs, non-profits.

    • Same kit as Team
    • Publish findings — we'll coordinate disclosure
    • Citation in quarterly security notes
    • * Subject to scope review
    Apply
start

Break in first.
Ship it cleaner.

Pull AdversAI today. Point it at an allowlisted target. See a trace in your terminal before your coffee lands.

Get the codeRe-watch the demo
· no card · no lock-in · MIT for the core ·