Threat Actors Could Misuse Code Assistant To Inject Backdoors and Generating Harmful Content

Researchers demonstrated an indirect prompt-injection vector against AI-driven coding assistants in which attackers seed public data sources or context attachments with malicious instructions. When the assistant ingests that tainted context, it can generate hidden backdoor code (example: a fetch_additional_data function contacting an attacker C2 and executing returned commands) that blends into legitimate code, bypasses moderation and code review, and can grant unauthorized remote access; the report urges stricter context validation and execution controls to mitigate the risk.