AI Red Team Scanner

Configure and run your scan

Nemesis sends adversarial prompts to your LLM using your own API key. Test a raw model, or paste your system prompt to test your configured AI deployment. This tool tests API-level responses only -- not your website or servers.

No data storedBrowser-only executionOWASP LLM Top 10 mapped
OK
What this tests: your LLM responses to adversarial prompts -- prompt injection, jailbreaks, data leakage, system prompt extraction, and more
x
What this does not test: your website, web servers, databases, network infrastructure, or application code
1

Target model

Select provider and model to test

e.g. gpt-4o, gpt-4-turbo, gpt-3.5-turbo

Stored in browser memory only. Never sent to Kyora IQ.

2

Attack modules

Select categories to test
3 of 16 modules · 29 tests
✓
Prompt Injection
LLM01:2025
Direct & indirect injection - instruction override, unicode obfuscation, delimiter escape, RAG pipeline, tool output injection.
11 tests
✓
Sensitive Data Disclosure
LLM02:2025
PII leakage from context windows, credential extraction, PHI inference, output-channel exfiltration via generated code.
12 tests
Data Poisoning
LLM04:2025
Tests whether injected malicious content in a simulated knowledge base influences model outputs and decisions.
6 tests
Improper Output Handling
LLM05:2025
Exfiltration-capable code generation, JavaScript cookie theft, markdown tracking pixels, DNS covert channels.
5 tests
✓
System Prompt Leakage
LLM07:2025
6 extraction vectors - direct ask, translation tricks, token completion, summarization, indirect inference, persona-based extraction.
6 tests
Vector & Embedding Weaknesses
LLM08:2025
RAG poisoning via malicious document chunks, indirect injection via retrieval context, cross-document instruction bleed.
5 tests
Misinformation & Hallucination
LLM09:2025
Tests whether the model can be pushed to assert false facts, amplify hallucinations, or generate targeted disinformation.
6 tests
Model Identity & Fingerprinting
LLM07:2025
Tests whether an attacker can determine which base model powers your product, what training data was used, and whether fine-tuning details can be extracted through systematic probing.
6 tests
Agentic Chain Attacks
LLM01:2025
Tests whether malicious instructions can propagate through multi-agent pipelines - agent A contaminating agent B, tool output hijacking, and cross-agent privilege escalation.
5 tests
Embedding & RAG Leakage
LLM08:2025
Tests whether RAG vector embeddings can be reverse-engineered to reconstruct source documents, cross-user context leakage, and the EchoLeak/Copilot attack vectors specifically.
6 tests
API Security
API:2023
Rate limiting, CORS misconfiguration, HTTP method enumeration, verbose error disclosure, missing auth headers, response metadata leakage.
6 tests
Injection Probing
LLM01:2025
SQL injection via chat interface, NoSQL injection, OS command injection, template injection, path traversal via prompts - the full McKinsey attack chain.
6 tests
Toxicity & Harmful Content
LLM09:2025
Hate speech generation, self-harm facilitation, targeted harassment, radicalization content, harmful medical advice, and dangerous instruction generation.
6 tests
Supply Chain Vulnerabilities
LLM03:2025
Third-party plugin hijacking, compromised model provenance, malicious tool substitution, dependency confusion attacks via plugin manifest injection.
5 tests
Excessive Agency
LLM06:2025
Agentic privilege escalation, unauthorized tool invocation, scope creep via instruction injection, autonomous action beyond intended permissions.
5 tests
Jailbreaks & Model Extraction
LLM10:2025
Persona hijacking, encoding-based bypasses, multi-turn escalation, model fingerprinting, training data memorization, resource exhaustion.
14 tests
3

Authorized use acknowledgment

Your API key will be used to run the selected tests and then discarded. Nothing is saved.

Scan summary

Modules selected3 / 16
Total tests29
ProviderOpenAI
ModelNot set
System promptNone
Est. duration~1 min

What happens next

1

Your browser connects directly to your provider's API using the key you entered — not via Kyora IQ's servers.

2

Each attack module fires its tests one by one. You'll see every prompt and response in real time with pass/fail results.

3

Each finding is scored by severity, mapped to OWASP LLM Top 10, and given a concrete remediation step.

4

Download your full report as HTML or PDF. Nothing is saved on Kyora IQ's end.

How we protect you

✓API key in browser memory only — never sent to Kyora IQ
✓Only official provider endpoints permitted
✓Scan results rendered locally — not persisted anywhere
✓Closing this tab discards everything immediately
✓No analytics, tracking, or data collection
⚠ Authorized use only

Only test models you own or have written permission to test. Unauthorized use may be illegal. Full policy →