Scott Thornton

Large language models remain vulnerable to jailbreak attacks, and single-layer defenses often trade security for usability. We present TRYLOCK, the first defense-in-depth architecture that combines four heterogeneous mechanisms across the inference stack: weight-level safety alignment via DPO, activation-level control via Representation Engineering (RepE) steering, adaptive steering strength selected by a lightweight sidecar classifier, and input canonicalization to neutralize encoding-based bypasses. On Mistral-7B-Instruct evaluated against a 249-prompt attack set spanning five attack families, TRYLOCK achieves 88.0% relative ASR reduction (46.5% to 5.6%), with each layer contributing unique coverage: RepE blocks 36% of attacks that bypass DPO alone, while canonicalization catches 14% of encoding attacks that evade both. We discover a non-monotonic steering phenomenon -- intermediate strength (alpha=1.0) degrades safety below baseline -- and provide mechanistic hypotheses explaining RepE-DPO interference. The adaptive sidecar reduces over-refusal from 60% to 48% while maintaining identical attack defense, demonstrating that security and usability need not be mutually exclusive. We release all components -- trained adapters, steering vectors, sidecar classifier, preference pairs, and complete evaluation methodology -- enabling full reproducibility.

AI-assisted code review is widely used to detect vulnerabilities before production release. Prior work shows that adversarial prompt manipulation can degrade large language model (LLM) performance in code generation. We test whether similar comment-based manipulation misleads LLMs during vulnerability detection. We build a 100-sample benchmark across Python, JavaScript, and Java, each paired with eight comment variants ranging from no comments to adversarial strategies such as authority spoofing and technical deception. Eight frontier models, five commercial and three open-source, are evaluated in 9,366 trials. Adversarial comments produce small, statistically non-significant effects on detection accuracy (McNemar exact p > 0.21; all 95 percent confidence intervals include zero). This holds for commercial models with 89 to 96 percent baseline detection and open-source models with 53 to 72 percent, despite large absolute performance gaps. Unlike generation settings where comment manipulation achieves high attack success, detection performance does not meaningfully degrade. More complex adversarial strategies offer no advantage over simple manipulative comments. We test four automated defenses across 4,646 additional trials (14,012 total). Static analysis cross-referencing performs best at 96.9 percent detection and recovers 47 percent of baseline misses. Comment stripping reduces detection for weaker models by removing helpful context. Failures concentrate on inherently difficult vulnerability classes, including race conditions, timing side channels, and complex authorization logic, rather than on adversarial comments.