Skip to content
DEV PREVIEW // PROTOTYPE
AI red teaming // evidence first

Put your AI
to the test.

Break it before someone else does.

Mallace probes how AI products fail and leaves a trail of evidence you can actually investigate.

Explore Mallace Read the docs
Prompt injection✕Jailbreaks✕Information extraction✕Guardrail bypass✕Instruction hierarchy✕Hallucination probing✕Bias✕Model drift✕Prompt injection✕Jailbreaks✕Information extraction✕Guardrail bypass✕Instruction hierarchy✕Hallucination probing✕Bias✕Model drift✕

One command

A full campaign.
Every result labeled.

Scope a target, fire adversarial cases, and get findings you can trace back to the exact request. Mock runs say they are mock. Failed runs stop instead of faking a pass.

  • Authorized targets only
  • Mock results always labeled
  • JSON · HTML · SARIF reports
OPERATOR@MALLACESIMULATED

ILLUSTRATIVE SESSION. SAMPLE DATA IN MOCK MODE. NOT A LIVE RESULT.

Confidence needs a paper trail

A convincing answer
isn’t proof of safety.

AI failures hide in the gap between what you intended and what the system actually does. Mallace pulls adversarial cases, responses, evaluation, and reporting into one inspectable workflow.

01

Probe the edges

Explore prompt injection, jailbreaks, information extraction, and other defined attack categories.

02

Preserve the evidence

Simulated results stay labeled. Failed execution and invalid judging never become confident findings.

03

Move toward remediation

Export findings for review. The proposed ForgeLoop bridge would connect approved findings to issues, PRs, and retesting.

The campaign

Four stages.
No shortcuts.

Every stage is inspectable. Nothing is silently substituted.

  1. 01

    Scope

    Define the authorized target, boundaries, and attack categories before anything fires.

  2. 02

    Attack

    Select and execute adversarial cases against the system under test.

  3. 03

    Observe

    Capture the actual response. Failed requests stop the run, never fake a pass.

  4. 04

    Evidence

    Export findings as JSON, HTML, or SARIF, each labeled with its execution mode.

Built in the open, locally

Know what’s real.
Know what’s next.

Mallace is an early prototype. The CLI and API exist today. A dependable live workflow, production access controls, and the operator dashboard are still being built.

Follow development
SYSTEM STATUS3/5 SHIPPED
Local CLI and prototype APIAvailable
Simulated campaign pipelineAvailable
JSON, HTML, and SARIF reportsAvailable
Live evidence and access controlsIn development
Operator dashboard and ForgeLoop bridgePlanned

Field notes

Beyond the demo.

View journal
Research note01

Why failed tests must never become successful evidence

NOV 4, 2026
Research note02

Find. Fix. Verify. The case for a closed remediation loop.

NOV 4, 2026

Run a simulated campaign locally

Find the failures.
Follow the evidence.

Get started $ python main.py --mock