Researchers Backdoor Open AI Model to Steal Credentials in Coding Agents

2 comments

• verdverm 2 hours ago
  1. these are older, weak models; newer or bigger models are more resistant to this

  2. don't give your agents access to credentials in the first place, build workflows around them instead

  example, I built a pr-review agent that has no credentials, no access to the internet, and no bash - it doesn't need these things to do a review
• skpd 2 hours ago
  [dead]