Summary

  • Researchers from Israeli firm Pillar Security have discovered a method of introducing silent backdoors into code generated by AI coding assistants.
  • The “rules file poisoning” attack involves the injection of malicious instructions into the AI training files, which could be assembled to produce malware.
  • Another element of the methodology is “unicode obfuscation”, where invisible characters in the files conceal harmful instructions that are recognisable to the AI models but invisible to human reviewers.
  • The third part of the process is “semantic hijacking”, where the AI is tricked into recommending pieces of code that bypass security best practices in order to introduce vulnerabilities.
  • Finally, because the malicious rule files can be present in repositories, the compromise can be persistent, spreading to future forks of the codebase.
  • The researchers have outlined a number of strategies to mitigate the risk, which include auditing AI configuration files for anomalies and validating the rule files as rigorously as executable code.

By Tal Eliyahu

Original Article