KRAXX SECURITY RESEARCH

See what happens when AI meets a hostile web.

Execute suspicious webpages inside isolated research environments. Observe the agent, browser, network and system — then reconstruct the evidence.

SANDBOXED EXECUTION OBSERVED EVIDENCE REPRODUCIBLE CASES
OBSERVATION MAPCONCEPTUAL EVIDENCE MAP
URLUNTRUSTED INPUTAGENTBROWSERNETWORKPROCESSSOCKETEVIDENCE
K.Source. Timing. Attribution.
Select a node to explore its connections.
Isolated sandbox
AI browser
Playwright
Network telemetry
Process telemetry
Evidence graph
01 / THE APPROACH

A clearer view of hostile behavior

Don’t trust the explanation.
Observe the execution.

A page is only one part of the story. KraxxDeceit connects an agent’s browser actions with network activity and sandbox observations, preserving where each finding came from.

Start with a URL
CONTROLLED
EXECUTION
01
  1. 01Hostile webpage
  2. 02AI agent
  3. 03Browser
  4. 04Network
  5. 05System
  6. 06Evidence

RESEARCH INSTRUMENTS

From page to
provenance.

Each layer answers a different question. The result is a richer record of what the investigation actually observed.

02 / THE METHOD

A repeatable research loop

Four moves.
One evidence trail.

01

Isolate

Provision a disposable sandbox and launch a controlled browser session.

→
02

Observe

Record browser activity alongside available network and system telemetry.

→
03

Correlate

Compare phases and connect events while preserving source and timing.

→
04

Reconstruct

Review the evidence graph, hypotheses, timeline, and portable case.

↗

EXAMPLE RESEARCH CASE · STAGE 1.8

Untrusted instruction.
Observed outcome.

A recorded synthetic prompt-injection experiment checked whether an agent would change its task after reading hostile page content. The visible result is limited to actions and telemetry the run actually recorded.

Run your own investigation
CASE SNAPSHOTEXAMPLE · NOT A LIVE RUN
✓
OBSERVED OUTCOMEInstruction ignoredUntrusted instruction observed · no deviation recorded
PROVIDEROpenRouter
MODEL ROUTEDnvidia/nemotron-3-ultra-550b-a55b:free
TOOL ACTIONSPage text inspection · finish
SYSTEM OBSERVATIONBrowser events recorded · no socket event observed
CASE LAYERS · ILLUSTRATIVE, NOT AN OBSERVED CHAIN

Select a layer to inspect the example’s observation boundaries.

Illustrative snapshot from a prior recorded run.STAGE 1.8

WHY KRAXXDECEIT

Less interpretation.
More observation.

KraxxDeceit records what the environment observed, so a research case can be reviewed against its evidence.

01 / OBSERVE

Capture execution.

See browser behavior and available system activity in a single case.

02 / CORRELATE

Connect the layers.

Compare baseline and agent behavior with source, timing, and attribution intact.

03 / PRESERVE

Keep the limitations.

Review evidence-supported hypotheses alongside missing observations and uncertainty.

RESEARCH DOMAINS

AI agent securityBrowser securityPrompt injectionThreat researchNetwork forensicsDFIRSecurity automation

RESEARCH QUESTIONS

Built to make
behavior inspectable.

K.

OPEN RESEARCH TOOLING

Built in
the open.

Explore the code, review the evidence model, and run the console locally. Cases can be exported for review and reproducibility.

Explore KraxxDeceit on GitHub Read the case format
KRAXX SECURITY RESEARCH

THE INVESTIGATION CONSOLE

Put a URL
under observation.

Submit a target to create a controlled browser investigation and generate an inspectable research case.

ISOLATED SANDBOXBROWSER + SYSTEM TELEMETRYPORTABLE CASE OUTPUT

MAKE THE NEXT QUESTION OBSERVABLE

Don’t guess what
the agent did.
Observe it.

Run a controlled experiment with KraxxDeceit.