§03Open Source & Research Contributions
OWASP GenAI Security Project
Multi-Technique Guardrail Bypass Evaluation
Five-stage evaluation sequence across four bypass technique families against a production chatbot. Demonstrates that patching individual bypasses fails, because the same intent gets through via different surface transformations.
OWASP GenAI Security Project
Multi-Turn Safety Bypass & System Role Override
Six-stage attack chain: control token injection, role-label spoofing, role-switching disclosure, Crescendo escalation. Mapped to LLM01, LLM02, LLM05, LLM07. Responsibly disclosed via MITRE.
Independent Research
Mechanistic Interpretability: Refusal in Open-Weight Models
Characterizing how refusal is represented in the residual stream of Qwen 2.5. Using targeted interventions to identify which directions are causally responsible for refusal vs. compliance.