Research
Smoke and Mirrors: Jailbreaking LLM-based Code Generation via Implicit Malicious Prompts | Proceedings of the 2026 IEEE/ACM 48th International Conference on Software Engineering
Publication date not yet evaluated · Added · 2 reports acm.org
Page published
Coverage timeline
Why it matters
CodeJailbreaker shows that instruction-following safety filters can be defeated by hiding malicious intent in auxiliary fields such as commit messages, meaning defenders of AI coding assistants must inspect implicit contextual cues rather than just explicit prompts.
CodeJailbreaker is a jailbreaking approach presented at ICSE 2026 that bypasses LLM safety mechanisms for code generation by keeping the visible instruction benign while encoding malicious intent implicitly in a covert channel, namely the commit message. Evaluated on the RMCBench benchmark, it substantially outperforms conventional jailbreaks that state malicious intent explicitly across three code generation tasks against models like GPT-4.