The UnboundCompute blog
Writing on web security
-
AI Agent Memory Poisoning: When a Planted Note Attacks Later
How ai agent memory poisoning works: one malicious note saved to an agent’s memory store sits quietly and hijacks future sessions long…
-
RAG Poisoning Attack: When Retrieved Documents Hijack the Model
A rag poisoning attack hides instructions in a knowledge base so a matching query pulls them into the model. See how it…
-
Code Interpreter Escape: Breaking Out of an AI Agent’s Sandbox
How a code interpreter sandbox escape works: injected text steers an AI agent’s generated code into reading secrets, exfiltrating data, and breaking…
-
Crescendo: The Multi Turn Jailbreak That Escalates Slowly
The crescendo multi turn jailbreak escalates over several benign turns until a model is led somewhere it would refuse outright. How it…
-
ANSI Escape Injection: Attacking AI Agents That Print to a Terminal
ANSI escape injection lets attacker controlled text rewrite what a CLI AI agent prints, so a human approves a destructive action they…
-
Multimodal Prompt Injection: Hiding Instructions in an Image
Multimodal prompt injection hides instructions inside an image or PDF so a vision model reads them as commands while text filters never…
-
Glitch Tokens: The Words That Break a Language Model
A glitch token attack feeds a language model rare, under trained tokens to break it. Learn how these dead tokens form, why…
-
Markdown Image Exfiltration: How a Chat UI Leaks Your Data
Markdown image exfiltration lets a chat assistant leak private data through an auto rendered image URL, no click needed. See how it…
-
LLM Backdoors: Hiding a Trigger in the Training Data
An llm backdoor attack plants a hidden trigger during training so a model looks aligned but flips on command. See how it…
-
Tool Output Injection: When an Agent’s Own Tools Lie to It
Tool output injection is when an AI agent trusts attacker controlled data from a tool and follows hidden instructions. How it works…