Cloudflare: Attackers are deceiving AI models with prompt injection

Cloudflare researchers found that seven AI models can be steered by prompt injection, especially when attackers hide lure comments inside code.

Small, subtle text cues sharply reduced malware-detection accuracy, while large repetitive prompts were more likely to trigger alarms.

When malicious instructions were buried in large code contexts, detection fell to very low levels, exposing a major weakness in context handling.

The study also found language-related bias in how models flag comments, suggesting AI security tools can be manipulated by both context and training-data bias.