The Risks of Code Assistant LLMs: Harmful Content, Misuse and Deception

In a detailed public assessment, Unit 42 examines the risks posed by code assistant LLMs used as IDE plugins, warning that both users and threat actors can abuse features such as chat, auto‑completion and unit test generation to inject backdoors, leak data or produce harmful content. The piece highlights indirect prompt injection through tainted external data sources and explains how context attachments can be hijacked when external files or URLs feed prompts that precede a user’s request.

It also covers direct model invocation risks, including the potential for threat actors to sell access to cloud‑hosted LLM services and for stolen session tokens to be misused in LLM-based attacks, with examples drawn from threats like LLMJacking. Mitigations focus on defensive practices such as thorough code reviews, careful examination of attached context, and enabling manual execution controls to prevent illicit actions.

Published on 15 September 2025 by Osher Jacob, according to Unit 42, the article urges developers and organisations to apply standard security measures across IDEs and model deployments to stay ahead of evolving exploitation techniques.