Skip to content
KageXKageX
Open Source · Browser Extension

Attack your AI like a real attacker would.

Mirage is a browser extension that red-teams AI systems from inside the browser. An autonomous agent runs the full recon → plan → test → judge → iterate loop, adapting to the target — with deterministic scoring that never trusts an LLM to grade itself.

Attacks across the whole kill chain

The agent reasons across attack classes mapped to the OWASP Top 10 for LLM Applications, grouped by the channel the payload rides in on.

Direct prompt
  • Prompt injection (LLM01)
  • System-prompt extraction (LLM06)
  • Multi-turn guardrail erosion
  • Structured-output / business-logic injection
  • LLM-as-judge bypass
Data-channel / indirect
  • Indirect prompt injection
  • RAG / knowledge-base poisoning
  • Membership inference
  • Markdown-image exfiltration
  • Multimodal caption injection
Agentic / tool
  • Confused-deputy tool auth (LLM07)
  • Insecure plugin / automation (LLM07)
  • Agentic data exfiltration (LLM08)
  • Tool-definition poisoning
  • Multi-agent trust-chain

Every model, as target or attacker

One provider abstraction drives the system under test and the attacker/judge brain alike.

Claude

claude-opus-4-8

OpenAI

gpt-4o

Grok (xAI)

grok-4.5

Local

ollama · mistral

How it works

01
Point it at a target

Attack the AI chat UI open in your active tab, or a direct API endpoint you configure — the same attack packs run against either.

02
Recon → plan → test → judge → iterate

The agent fingerprints the target, forms a hypothesis, attacks (multi-turn), reads the result, and adapts — scoring on real, decoded side-effects, not an LLM's opinion.

03
Read the findings

Wins stream into the side panel with full transcripts and an OWASP-LLM mapping. Export a shareable HTML or JSON report.

Adaptive agent

Recon → plan → test → judge → iterate

Deterministic scoring

Wins require a real decoded side-effect

Exportable reports

HTML + JSON with full transcripts

Multi-provider

Claude, OpenAI, Grok, and local models

Red-team your AI in minutes

Free and open source. Install the extension, point it at your AI, and see what breaks — responsibly, on systems you're authorized to test.