Break it before they do.

An autonomous pentester that maps your attack surface, proves the exploit, and opens the fix as a pull request.

Get started Talk to us
run your first engagement from the CLI →
Research record

Built in real attack surfaces, not a slide deck.

Rokkhe was used in authorized security research across aviation, mobility, secure messaging, commerce, delivery, and fintech. The platform keeps the same discipline: a finding is not promoted until its evidence is reproducible.

OWASP
MITRE ATT&CK
CVE / NVD
CISA KEV
ExploitDB
Semgrep
Nuclei
Trivy

Techniques and signatures sourced from published security research

A finding without a working exploit is a guess with a severity label on it.
Rokkhe engineering principleWhy every run ships its proof
The engine

How an engagement runs

Rokkhe builds a working map of the target, sends agents to break it inside a sandbox, and keeps whatever they manage to prove.

How it works
Three stages, one run
STAGE 01

Maps the surface

Routes, parameters, auth boundaries, and dependencies — enumerated the way an adversary would.

STAGE 02

Proves what is real

Parallel agents write and run exploits in a container sandbox, keeping the trace that confirms each one.

patch ready
STAGE 03

Opens the fix

The remediation lands as a reviewable diff with its evidence attached, and stays in sync with the finding.

What it actually catches

Missing authorization, tenant leakage, and logic flaws that no pattern match will ever see.

Three findings from a single run

Invoice export

CRITICALSeverity
42,113Records
1,204Tenants
▤ app/api/invoices/export.py
01 −
01 +
02 +
Rokkhe

Authorization applied after the query

See the proof →

Webhook secrets

HIGHSeverity
318Objects
IDORClass
▤ app/api/webhooks/[id].ts
01 −
01 +
02 +
Rokkhe

Path parameter trusted as identity

See the proof →

Rate limiter

MEDIUMSeverity
HeaderVector
UnboundedImpact
▤ api/rate_limit.py
01 −
01 +
02 +
Rokkhe

Limiter keyed on a spoofable header

See the proof →
Browse more findings
Agentic security operations

Run the SOC as one visible, governed evidence loop.

Bring defensive signals and offensive proof into the same system. Every plan, hypothesis, approval, action, and outcome stays inspectable.

Detect + normalize

One queue for signals that still retains where every fact came from.

Ingest alerts from security tools, quarantine malformed events, deduplicate safely, and preserve the original evidence under the correct workspace.

Source provenance and replay state retained
Alert queueWorkflow preview
HIGHIDENTITYImpossible travel + new OAuth grant evidence retained
CRITENDPOINTUnsigned binary spawned a shell evidence retained
HIGHCLOUDPublic policy change on sensitive store evidence retained
normalizededuplicateworkspace scopequarantine + replay
Case RK-SOC-104Analyst review
Alertsource event
Identityservice account
Assetpayments-api
Hypothesisevidence-backed
Decisionreview required
cited evidence approval boundary
Triage + correlate

Agents investigate with context—not a hidden chain of guesses.

Retrieve related assets, identities, findings, and artifacts; compare hypotheses; expose confidence and evidence; then hand the decision to an analyst.

Plans, evaluations, and review decisions stay visible
Durable by design

Absorb security-event bursts without losing ordering, ownership, or evidence.

Bounded batches, backpressure, deduplication, leases, retries, dead-letter handling, and replay keep the pipeline recoverable instead of process-local.

01Receive
02Normalize
03Correlate
04Durable job
05Case + evidence
bounded batches · backpressure · deduplication · replay
Detect as code

Generate, test, translate, deploy, and reconcile detections.

A proposal stays a proposal until fixtures, target controls, and required reviewers accept it.

detections/suspicious-oauth.ymlProposed
title: Suspicious OAuth grant after impossible travel
logsource:
  product: identity
detection:
  selection: new_grant
  condition: selection and risk_context
status: test_required
syntax fixtures○ human approval
Respond with supervision

Bound every response with policy, approval, and outcome verification.

Consequential actions are scoped, idempotent, reversible where possible, and independently checked.

Response planApproval required
1
Disable compromised tokenreversible · scoped identity
2
Preserve session evidenceappend-only artifact
3
Verify access is revokedindependent outcome check
Policy fence activeNo action runs before authorization and reviewer approval.
Native security telemetry adapters, normalized into one evidence loop
From SOC to cloud

One evidence model across the security stack.

Native normalization preserves source context, turns correlated signals into cases, and keeps the analyst’s review boundary visible. Every connector stays explicit about its health and configuration state.

SplunkES · HEC
ElasticELK · ECS
AWSSecurity Hub · GuardDuty
Google CloudSecurity Command Center
Microsoft Sentinelincidents · Defender
CrowdStrikeFalcon detections
Wazuhalerts
Private model deploymentoperator controlled

Security reasoning that can stay inside your boundary.

Our deployment profile combines 284B, 13B, and 1M-class security models with FP4 + FP8 mixed precision. Kimi K3 can be deployed on-premises for organizations that require a private inference path.

On-prem inferencePrivate data pathProvider-configured
Model selection, capacity, and residency are defined in the deployment plan—not assumed by a marketing tier.
Delivery systems

Proof should arrive where work gets done.

Rokkhe keeps code review, security checks, ticketing, and notifications tied to the same finding and workspace boundary.

GitHubpull requests · checks
GitLabmerge requests · pipelines
Jiravalidated issue routing
Linearremediation sync
Slackcase alerts · escalation
Discordrouted notifications
Rokkhe CLI

A real operator interface, not another browser tab.

Use the CLI when the browser is not the right surface: authenticate from a terminal, preflight a deployment, start an authorized engagement, inspect retained evidence, or hand off a verified finding.

Python 3.12+ · macOS and LinuxUser-level install · no root accessRelease checksum verified by installer
Read the CLI reference
cli.rokkhe.com · release channelv1.2.0
01 Install02 Verify03 Connect
One-line install
$ curl --proto '=https' --tlsv1.2 -fsSL https://rokkhe.com/install.sh | sh
installer → checksum → isolated environment✓ SHA-256 checksum verifiedInstalled: ~/.local/bin/rokkhe$ rokkhe doctorPreflight passed. Ready to connect to your Rokkhe deployment.

Inspect first, if you prefer. Download the installer, review it, then run it locally. The command is selectable even when clipboard access is blocked.

See it work

What Rokkhe finds, how it proves it, how it fixes it.

Plays on its own · pick any step
new-scan · scopeSetup
Repositoryacme/checkout-apimain · 41k LOC · Python
Live targetstaging.acme.devsandboxed copy
Scope contract
Only the targets you name
Agents and tools run in a container sandbox
Your model key, your account, your spend
Every action written to the run log
illustrative walkthrough
Scope and control

It only touches what you name

Every engagement runs against a boundary you write, and every action it takes is recorded against that boundary.

Scope contract

Targets Rokkhe is permitted to reach.

1Repositoryacme/checkout-api
2Live targetstaging.acme.dev
3Never*.prod.acme.dev
Boundary

Your targets, your keys

Name the repositories and hosts in scope. Agents and tools run in a container sandbox on infrastructure you operate, against model credentials you configure.

Read the scope model
Containment heldawaiting analyst approval
Authority

It proposes, you decide

Consequential response stops at an approval gate carrying the plan, the evidence, and the reasoning. An unverified proposal is never shown as a confirmed result.

See the audit trail
Your workflow

Where the fix lands

Findings become reviewable work in the systems your team already opens every morning.

All integrations
Four ways Rokkhe reaches your team
01

Pull requests

The remediation arrives as a diff against the branch you nominate, evidence attached.

rokkhe mcp
02

Model context

Expose finding context and evidence to whichever coding agent your team drives.

03

CLI and CI

Start an engagement per release, per branch, or on a schedule from your pipeline.

Slack · Jira · Linear
04

Triage channels

Run state, confirmed exploits, and approval requests routed to the owning team.

Sovereign by design

Built for teams whose code and findings cannot leave the environment that produced them.

Deployment guide
01

Self-hosted

Run the entire platform inside your own cloud or air-gapped network.

02

Private inference

Configure a private model path so prompts and source never reach a third party.

03

Workspace isolation

Every record carries its boundary, enforced in the repository layer rather than the UI.

FAQ

Questions, answered.

Can't find your answer here? Get in touch.

Find out what is exploitable before someone else does.

Get started View pricing