Offensive security research lab

We find the paths
scanners miss.

Finding alerts is easy. Proving what is actually exploitable is the research. RootXLabs discovers and validates real attack paths across applications, AI systems, cloud infrastructure, open-source software and emerging technology.

CVEs assigned
5 CVEs assigned
Vendors that acknowledged findings
8 vendors acknowledged
Bugcrowd global rank
Bugcrowd #700 globally

Research Exploit validation Attack-path analysis

attack-path/01entry → impact
ENTRYInternetSURFACEAPI gatewayTRUST BOUNDARYAuthenticationINTERNALInternal serviceESCALATIONPrivileged workerIMPACTSensitive dataRepositoryAI agentCloud

[ How we think ]

Finding patterns is easy.
Understanding the exploit chain is the research.

A scanner stops at the signal. We keep going — through the context it sits in, the trust boundary it crosses, and the chain that turns a handful of individually weak behaviours into something with real impact. Only the last step is worth an engineer's attention.

  1. 1

    Signal

    A scanner alert, an odd response, an assumption in the docs. On its own it means nothing.

  2. 2

    Context

    Where does it sit? What calls it, what trusts it, what runs with more privilege than it should.

  3. 3

    Trust boundary

    The line the design assumes an attacker cannot cross. Most real bugs live in how that line is enforced.

  4. 4

    Exploit chain

    Weak behaviours, linked. One reachable from the next, until the sequence reaches something that matters.

  5. 5

    Real impact

    Data read, identity assumed, control gained — reproduced safely, with the exact steps to reproduce it.

[ Capabilities ]

Research across the modern attack surface

Five domains, one method: understand the architecture, find the trust boundary, and prove the path across it. Each surface breaks differently — the reasoning that reaches the bug does not.

01appsec

Application security

Complex web apps, APIs, authentication and the business logic underneath — where a request that is individually valid becomes an exploit in sequence.

Request
Authorization
Backend
Privileged action
  • Authentication & session logic
  • Authorization & tenant isolation
  • Business-logic abuse
  • API & GraphQL surface
02ai

AI security

LLM applications and agentic systems, where the trust boundary moves at runtime — a model decides which tool to call, and the tool reaches an internal system.

User
Model
Tool call
Internal system
03cloud

Cloud security

IAM, service boundaries and identity chains — how a single over-scoped role or a reachable metadata endpoint turns one foothold into everything.

Identity
Assume role
Metadata
Secrets
04oss

Open-source security

Source audits, dependency chains and package ecosystems — reading the fix as carefully as the bug, because the patch is where the next variant hides.

Repository
Dependency
Sink
Advisory
05emerging

Emerging technology

Browsers, desktop apps, devices and firmware — new attack surfaces where the trust model is still being written, and the interesting bug is architectural.

Untrusted input
Parser
Sandbox
Host

[ Research methodology ]

From attack surface to proven impact.

The same five moves on every engagement, whatever the surface. No step is skipped, because the interesting bug almost always lives in the one a checklist would skip.

01

Map

Understand the architecture, assets, identities and exposed surface — the system as it actually is, not as the diagram claims.

02

Reason

Identify the trust assumptions, authorization boundaries and behaviours that are individually odd but not yet exploitable.

03

Chain

Connect weak behaviours into a path an attacker could actually walk — each step reachable from the one before it.

04

Validate

Reproduce the issue safely on a clean build, with a negative control proving the vulnerable path is the one that fired.

05

Report

Deliver a reproduction a triager can re-run cold, the realistic impact, and the remediation — no CVSS theatre.

[ Signal over noise ]

Less noise.
More proof.

A scanner hands you a queue and calls it a result. We hand you the two or three chains that actually reach something, each one walked end to end. The rest was never the work.

Typical scanner rununsorted
120+findings, mostly unreachable

Every dot is someone's afternoon: open it, decide if it's reachable, decide if it matters. Most are neither.

What RootXLabs hands backvalidated
3validated attack paths
  1. 1
    Public endpoint
    internet-facing
  2. 2
    Authorization gap
    trust boundary
  3. 3
    Internal service
    not meant to be reachable
  4. 4
    Privilege boundary
    role escalation
  5. 5
    Sensitive operation
    proven impact
Proven impact

[ Security research ]

Research that survives real-world validation.

Findings reported by RootXLabs and confirmed by the vendor. Published advisories link to their CVE record; work still under coordinated disclosure is listed only once a fix has shipped.

RXL-2026-002

CVSS 7.4

Microsoft

PowerShell

Published advisory. Technical detail is on the linked CVE record.

Information disclosureCVE-2026-58612
Published · credited

RXL-2026-001

CVSS 7.8

Microsoft

PowerShell

Published advisory. Technical detail is on the linked CVE record.

Remote code executionCVE-2026-40400
Published · acknowledged

RXL-2026-003

High

AXIS Communications

Camera Station

Advisory details withheld by the reporter.

Details withheldCVE-2026-63283
Published · details withheld

RXL-2026-004

CVSS 6.5

Apple

WebKit

Published advisory. Technical detail is on the linked CVE record.

Information disclosureCVE-2026-64778
Published · credited

RXL-2026-005

CVSS 8.8

Microsoft

PowerShell

Published advisory. Technical detail is on the linked CVE record.

Remote code executionCVE-2026-70337
Published · acknowledged

Findings acknowledged by

  • Google
  • Apple
  • Meta
  • Microsoft
  • Cisco
  • Perplexity
  • Dell
  • NASA
The full record

Coordinated disclosure on every finding · No detail published that would help reach an unfixed system.

[ Hall of Fame ]

Research acknowledged by the organizations that ship it

Every logo below is an organization that independently reproduced a RootXLabs finding and credited it — the calibre of software our research holds up against. Not customers, not endorsements: acknowledgements, which are not ours to award.

Google
Apple
Microsoft
Meta
NASA
Cisco
Dell
OpenAI
Google
Apple
Microsoft
Meta
NASA
Cisco
Dell
OpenAI
Atlassian
Under Armour
Volkswagen
Perplexity
eToro
Arlo
Thales
Kistler
Atlassian
Under Armour
Volkswagen
Perplexity
eToro
Arlo
Thales
Kistler
8+ vendors acknowledged5 CVEs assignedBugcrowd #700 globally

[ Research areas ]

The classes we go deep on

Not a taxonomy for its own sake — these are the bug classes where the gap between “a scanner flagged it” and “it is exploitable” is widest, and where careful reasoning pays off.

Authentication
Authorization
Sandbox escape
Remote code execution
SSRF
Supply chain
AI agents
Cloud IAM
Browsers
Desktop apps
Firmware
API security
Business logic
Deserialization
Path traversal
Privilege escalation

[ Attack-path demo ]

One weakness rarely tells the whole story.

A mock environment. Walk the chain — each node is a step an attacker takes, and the same node reads differently depending on where you are standing. Hover, tap, or step through it.

environment mapnode 1 / 5
Step 1Entry

Public API

edge · internet-facing

Entry point
An unauthenticated endpoint that echoes a request parameter into a downstream call.
Observation
The gateway forwards a client-supplied identifier without re-checking who the caller is.
Trust assumption
“The internal service trusts anything that reached it through the gateway.”
Exploit primitive
Request forgery into the internal network path.
Impact
Foothold: the attacker can now speak to services never meant to be reachable from outside.

[ Who we work with ]

When tooling reaches its limit, deeper reasoning begins

The teams we do our best work with have already automated the easy part. What is left is the reasoning a scanner cannot do — and that is the whole job.

Engineering teams

Shipping fast on a surface that outgrew the last review. You want to know what an attacker reaches, not a list of linter warnings.

Security teams

Already running the scanners. You need a second set of hands for the exploit chains automation cannot reason about.

AI companies

Models wired to tools and internal systems. The trust boundary moved at runtime and the old threat model no longer covers it.

Cloud-native platforms

Identity chains, service meshes and metadata endpoints, where one over-scoped role is the whole attack path.

Open-source maintainers

A patch that needs a second reader, or a report that arrived and you want it validated before you act on it.

Product security teams

Browsers, desktop apps, firmware — surfaces where the interesting bug is architectural and needs a manual eye.

[ Engagements ]

Three ways to put us on your surface

Whether you need one hard question answered or a standing adversary who tracks your surface as it changes, the method is the same and the output is proof.

A scope, a timebox, a proof.

Offensive security assessment

We take a defined target and attack it the way a motivated adversary would — architecture first, then the chains. You get validated attack paths and reproductions, not a scanner export.

  • Web, API, cloud, AI or native scope
  • Attack-path reasoning, not checklists
  • Every finding reproduced safely
  • Remediation guidance that fits your stack
Discuss a target
A hard question, answered.Most engagements

Security research

Deep research against a specific product, protocol or dependency — source audit, patch analysis, variant hunting — carried through to a working proof and coordinated disclosure.

  • White-box source review
  • Patch-bypass & variant analysis
  • Working proof-of-concept
  • Disclosure handled end to end
Discuss a target
A standing adversary.

Continuous research partnership

An ongoing relationship where we track your surface as it changes — new services, new models, new dependencies — and keep pressure-testing the boundaries between them.

  • Recurring, scoped engagements
  • Coverage that follows your roadmap
  • Named point of contact
  • Findings routed where your team works
Discuss a target

[ Responsible research ]

Credibility, not certification theatre

We hold access to systems that would hurt if we were careless with them. These are the constraints the work runs under — no badges we have not earned.

Coordinated disclosure

The vendor hears first, always. Nothing is published that would help someone reach a system still running the vulnerable version.

Evidence-driven findings

A finding is not a finding until it reproduces. If we cannot prove it, it does not leave the lab.

Manual validation

Automation reads everything; a human decides when a story is too neat. The exploit is stood up by hand.

Clear reproduction

Every report includes steps a triager can re-run cold, with a negative control showing which path actually fired.

Impact-focused reporting

Severity is what the impact supports — not what the report would be worth if it were higher.

Secure handling of research data

Scope is sacred and access is least-privilege. We take the minimum needed to prove it, and drop it when the work is done.

[ Research notes ]

What we can show you: the method

The findings are under coordinated disclosure. The way they were produced is not — and it is the part worth judging us on.

All notes

Work with us

Have a system worth breaking before attackers do?

Tell us what you want us to look at. We will tell you honestly whether it is the kind of target our research reaches — and if it is, what a first engagement would prove.

Scoped & authorized testing only · Coordinated disclosure on every finding