Researchers Highlight How Poisoned LLMs Can Suggest Vulnerable Code

Researchers presented CodeBreaker, a technique for poisoning code-completion model training data to cause LLMs to suggest vulnerable or backdoored code that evades static analysis; the paper builds on prior methods (COVERT, TrojanPuzzle) and warns that developers and model creators must better vet training data and review AI-suggested code to avoid introducing exploitable code into software.