ShadowProbe AI
Understand what your prompts
are really doing
Not every prompt is harmless. ShadowProbe AI exposes hidden risks in user inputs - from prompt injection to adversarial intent to sensitive data leakage - so you can catch issues before they escalate.This system explores how adversarial prompts are detected - not all attacks are obvious, and some will get through.
Simulate → Detect → Analyze
red agent
Simulates adversarial behaviorGenerates or ingests prompts designed to test system boundaries, including injection attempts and sensitive queries.Acts as the attacker in your testing pipeline.
blue agent
Detects and classifies threatsIdentifies known attack patterns and maps them to MITRE ATT&CK techniques for structured analysis.Transforms raw prompts into actionable security signals.
llm analysis
Context-aware risk analysis.Goes beyond rule-based detection by evaluating prompt intent,
language patterns, and potential attack vectors.Flags nuanced threats that traditional filters miss.
real-time threat
detection
ShadowProbe AI simulates adversarial prompts and analyzes them using a combination of rule-based detection and LLM reasoning to uncover hidden risks before they escalate.Identifies and classifies threats such as:
○ Credential harvesting attempts
○ Prompt injection and system manipulation
○ Sensitive data exposure risks
○ Social engineering and behavioral signalsGiving you visibility into risks that traditional filters often miss.
our capabilities
ShadowProbe combines detection, analysis, and context-aware reasoning to surface risks that traditional filters miss.
LLM-powered reasoning
Evaluates context and intent to uncover subtle threats beyondrule-based detection.Real-time risk scoring
Instantly classifies prompts as LOW, MEDIUM, or HIGH risk based on detected patterns and intent.MITRE ATT&CK mapping
Aligns detected behaviors with known adversarial techniques for structured security analysis.Memory-based trend tracking
Tracks patterns across runs to identify recurring risks and evolving behaviors.Interactive security dashboard
Visualizes results, risk distribution, and recent activity in real time.
the risk
why it matters
Prompts are becoming a new attack surface.As LLMs are integrated into real systems, adversarial inputs can manipulate behavior, expose data, and evade detection.ShadowProbe helps you see these risks before they escalate.
Security doesn’t stop at the model — it starts at the prompt.
ADVERSARIAL AI · PROMPT DEFENSE
Amie Twyford · ShadowProbe AI
I built ShadowProbe AI to explore how adversarial prompts and prompt injection attacks can be detected in real time. The system combines rule-based detection, MITRE ATT&CK mapping, and LLM-driven analysis to surface risks like credential exposure, social engineering, and sensitive data leakage.I’m especially interested in how AI systems behave under pressure, and how we can design tools that make those behaviors more visible, interpretable, and secure.Focused on opportunities in AI security and applied machine learning.
explore the system
test shadowprobe
Run prompts, simulate adversarial behavior, and see how ShadowProbe detects and explains risk in real time.Built as an interactive demo to showcase how AI systems can be tested, understood, and secured. // ready for next input
© All Rights Reserved. Amie Twyford | New York City