Yazan Boshmaf

Code LLMs can be reliably poisoned to generate vulnerable code by subtly manipulating the style of the input prompt, even while maintaining strong performance on standard benchmarks.