Published work on AI safety, adversarial robustness, and machine learning systems.
AI-safety guardrails that gate agent actions behind the Lean 4 kernel can be broken with documented features alone; a 33-trial study across two codebases and three Claude tiers found models wrote kernel-inconsistent code 0/12 under open optimization but 6/6 once the divergence was specified under a performance cover story.
Unsupervised attack detection via cross-perturbation generation profiling — 0.94–1.00 AUROC across six LLMs (8B–32B) with zero training data, where token-level baselines plateau at 0.56–0.83.
Nucleotide virulence factor benchmarks are inflated by ~0.30 AUROC from organism confounds and gene-family leakage.
Codon frequency + logistic regression matches deep models genus-to-genus once you fix the data leakage that inflated every prior benchmark by ~0.30 AUROC.