malicious-code

This paper reveals that grammar-constrained decoding (GCD) can be exploited as a jailbreak attack (CodeSpear) to induce LLMs to generate malicious code, and proposes a defense (CodeShield) that preserves safety under such attacks.

Johannes Link, maintainer of the Java library jqwik, added malicious prompt injection to disrupt AI usage of the library, sparking debate on AI ethics and open-source maintainer rights.