Loading...
Selected work
Open-source benchmark proving traditional SAST tools score near 0% on agentic code. 57 annotated cases across 4 languages with real CVEs, PR simulation mode, and pluggable scanner adapters, mapped to OWASP Agentic Top 10.
PyPI package with GitHub Actions integration. Uses 3 AI agents in parallel blind review with an adversarial debate engine for comprehensive code security analysis.
Meta-agent pentesting orchestrator with headless browser automation, confidence scoring, and 73% accuracy on XBOW benchmarks.
Architecture-aware security scanner combining SAST and DAST agents with support for 11+ programming languages.
Python vulnerability analysis agent integrating static analyzers (Bandit, Semgrep) with a RAG layer to produce vulnerability reports and remediation guidelines without model fine-tuning, improving detection accuracy by 35%.
Contributed the BadCharacters probe to jailbreak LLMs with adversarial perturbations (zero-width spaces, homoglyphs, BiDi overrides, and backspace deletion), using differential evolution and Levenshtein distance optimization for adversarial LLM testing.
Identified and fixed SQL injection in the Hive metastore bookkeeping and Presto partition predicates via identifier allow-listing (PR #66751).