Miller Engelbrecht -- Security Research

Writing

Clone This Repo and I Own Your Machine ↗

May 21, 2026 ai-security bug-bounty

0din.ai Submissions

Guardrail Jailbreaks via Contextual Reframing

20+ submissions

Content safety filter bypasses across multiple frontier AI models using prompt reframing tactics: academic chemistry, forensic training scenarios, pharmaceutical analysis, verbose simulation, patent litigation framing, and incremental escalation patterns. Each variant bypasses model-level guardrails without triggering refusal heuristics.

ai-safety multiple models 0din.ai

Indirect Prompt Injection via Embedded Content

multiple submissions

Data exfiltration through indirect prompt injection in AI-powered applications that process external content: email clients, document readers, and web browsing agents. Injected instructions embedded in legitimate-looking content redirect agent behavior to leak conversation context or user data.

prompt-injection exfiltration 0din.ai

Agentic Attack Surfaces in AI Coding Tools

multiple submissions

Attack chains targeting AI coding agents through workspace configuration files, package manager lifecycle hooks, and git repository metadata. Includes DNS-controlled payload delivery via npm postinstall, invisible Unicode injection in commit messages, and AGENTS.md trust exploitation.

ai-security coding-agents 0din.ai