Security framework for AI agents
Open-source middleware that detects and defends against AI Agent Traps. Scan inputs, RAG chunks, and outputs before they reach your LLM.
What it detects
Content Injection
Hidden HTML, metadata injection, dynamic cloaking, syntactic masking
Behavioural Control
Jailbreak patterns, data exfiltration, unauthorized sub-agent spawning
Cognitive State
RAG poisoning, memory poisoning, contextual learning manipulation
Semantic Manipulation
Biased framing, oversight evasion, persona hyperstition
Encoded Payloads
Instructions written in base64, hex or rot13. They are decoded and rescanned, and reported only when the decoded text is itself an attack, so images, JWTs and hashes stay clean
ASCII Smuggling
Instructions re-encoded into invisible Unicode so they render as nothing to a reviewer but are still read by the model — detected and stripped
Credential Exposure
Leaked API keys, tokens, private keys, and connection strings in ingested content — redacted in evidence so a scan never leaks what it finds
Dependency Substitution
Package installs that skip the trusted registry: custom index flags and config commands, installs from a raw URL, and a resolver redirected to a proxy (strict only)
Tool-Call Tampering
A download piped into a shell, a shell run on a downloaded script, or decoded text run by a shell, inside the command of a tool call. The same line as plain text is not flagged: install guides write it too
Response Anomaly
Scores commands and tool call arguments by a download that is run (inside a tool call), secrets sent out, layered encoding, credentials in arguments and string entropy. One signal reports at strict, two or more at balanced
Multi-turn / Session
Cross-turn split-payload detection via scanSession — instructions chopped across conversation turns
ML Classifier
DeBERTa-v3 multi-label model via ONNX. Optional async detection alongside regex patterns.
Systemic + HITL
Congestion, cascades, collusion, approval fatigue, social engineering
Quick start
import { AgentArmor } from '@stylusnexus/agentarmor';
const armor = new AgentArmor();
// Scan any text before it reaches your LLM
const result = armor.scanSync(userInput);
if (result.threats.length > 0) {
console.log('Threats detected:', result.threats);
const safe = armor.sanitize(userInput, result);
}
// Filter RAG chunks before context assembly
const clean = armor.scanRAGChunksSync(chunks)
.filter(r => r.threats.length === 0);
Eval results (v0.6.0 patterns)
| Strictness | Detection Rate | False Positive Rate |
|---|---|---|
| Permissive | 82.1% | 0.0% |
| Balanced | 91.0% | 0.0% |
| Strict | 91.0% | 0.0% |
105 curated samples (67 adversarial, 38 benign) from WASP, HackAPrompt, Greshake et al., and 2025-2026 real-world incidents.
Includes 10 samples from real-world attacks (MCP poisoning, RAG saturation, supply chain injection) that regex does not yet catch, measuring the gap the ML classifier closes. On the original 49 adversarial samples, regex detection is 100% at balanced.