AI Security Research Lab
Exploit-oriented research for agentic AI systems. Deterministic harnesses, reproducible findings, and open methodology for prompt injection, tool boundary attacks, and agent vulnerability disclosure.
Research Capabilities
Prompt Injection
Deterministic evaluation of prompt injection as an exploitable attack vector across RAG, browser, and tool-runner agent patterns.
Evaluation Harnesses
Reproducible testing frameworks that score agent security with binary exploit/no-exploit results — no LLM-as-judge.
Vulnerability Disclosure
Responsible disclosure program for AI agent platform vulnerabilities — CVE coordination and vendor notification.
Tool Boundary Analysis
Mapping the security boundary where LLM reasoning meets tool execution — exfiltration, policy override, and unsafe invocation paths.
Research
Exploit-oriented security research for agentic AI systems. Reproducible tools, deterministic findings, open methodology.
Prompt Injection Still Becomes a Security Bug at the Tool Boundary
We built a deterministic evaluation harness that models three common agent patterns — RAG assistants, browser agents, and internal tool runners — and measures prompt injection as a concrete exploitation problem: exfiltration, policy override, and unsafe tool invocation. 9 scenarios. 9 exploit successes. Binary scoring. No LLM-as-judge.
Harness MVP
Prompt Injection Evaluator
Deterministic harness that scores prompt injection across 9 agent scenarios. Binary exploit/no-exploit results with reproducible methodology.
RAG Pattern Tests
3 attack vectors against retrieval-augmented generation agents — exfiltration, policy override, and unsafe tool invocation.
Browser Agent Tests
3 attack vectors against browser-based agents — DOM injection, cross-origin exfiltration, and privilege escalation via tool chaining.
Vulnerabilities
Responsible disclosure for AI agent platform vulnerabilities. Findings coordinated with vendors through CVE assignment where applicable.
Prompt Injection Exploitation via Tool Boundary
Deterministic evaluation across 9 agent scenarios (RAG, Browser, Tool Runner) showed 100% exploit success rate for prompt injection attacks targeting exfiltration, policy override, and unsafe tool invocation. No LLM-as-judge — binary scoring with reproducible methodology.
Report a Vulnerability
CyberLab coordinates responsible disclosure for AI agent platform vulnerabilities. If you've discovered a security issue in an agentic AI system, contact our disclosure program.
Open Research for Open Systems
Our findings are reproducible, our methodology is transparent, and our tools are designed to be understood — not trusted on authority.