Software engineer with deep knowledge of AI security and AI research. I build and study systems that keep machine learning models, and the software around them, trustworthy.
Jailbreaks and adversarial attacks on large language models, and practical defenses such as perplexity-based detection of adversarial suffixes.
Measuring how models behave under attack, with reproducible experiments on benchmarks like JailbreakBench.
Years of building large-scale infrastructure, test automation, and software supply chain integrity tooling.
How adversarial suffixes such as GCG trick safety-trained models, and the design choices behind a perplexity-based detector that reaches 100% recall on GCG suffixes at about 1% false positives on JailbreakBench.