OpenAI addressed ShadowLeak in August, but has been quiet about exactly how, suggesting that ShadowLeak and tricks like it may yet still have legs. ShadowLeak: Non-Detectable Email TheftWhat happens, exactly, when you ask ChatGPT to do something with your emails? Addressing Indirect Prompt Injection Attacks on AI AgentsRadware reported the ShadowLeak vulnerability to OpenAI in June. Organizations can try to compensate with their own security controls — for example, by vetting incoming emails with their own tools. In his opinion, the only way to effectively address threats geared to AI agents is by employing yet more AI tools.