Beware What You Autocomplete: Forensic Attribution of Backdoored Code Completions

Researchers have demonstrated that AI-driven code completion tools are susceptible to backdoor attacks where malicious training data causes the models to suggest insecure code segments. This vulnerability poses a risk to software security, as developers relying on these systems might inadvertently introduce exploitable flaws into their applications.