Is indirect prompt injection still a big threat as models get more advanced? (realarcherl.github.io)

Latest research indicates that the threat of indirect prompt injection attacks on advanced AI models is diminishing, particularly against the newest iterations from leading labs like Anthropic and OpenAI. In comparative tests, models like Claude Opus 5.5 and GPT-6 Astra showed significantly reduced attack success rates—1% and 8.5% respectively—compared to older models which were more vulnerable. However, attackers have shifted focus to exploiting weaknesses in older models and untested pathways in newer ones, showing that while advancements have been made, the threat persists, especially from out-of-lab attackers employing crude tactics. These developments underscore the importance of rigorous testing before deploying AI models in real-world applications. Labs are increasingly publishing "system cards" detailing how their models were trained and tested, incorporating insights from external red-team companies like Gray Swan. This transparency aims to bolster model robustness against attacks. Nevertheless, the ongoing risks from indirect injections through tools and APIs highlight a need for continuous improvement and vigilance, particularly regarding the paths through which models can be unknowingly manipulated.