Wait, so what is the vulnerability precisely? Article says it requires prompt injection. Doesn't prompt injection open up a whole range of vulnerabilities though? How is the prompt being injected in this scenario? Does the attacker basically assume root access to the target's machine?
edit: Ah, I see - they created a custom GPT (as in GPTs) that enables the attack explicitly. I suppose that's problematic, but then so is the inability to inspect a custom GPT's system prompts. If OpenAI is really concerned about this they should have an approval process similar to the App Store. Or, you know - make the whole situation easier to inspect.
edit: Ah, I see - they created a custom GPT (as in GPTs) that enables the attack explicitly. I suppose that's problematic, but then so is the inability to inspect a custom GPT's system prompts. If OpenAI is really concerned about this they should have an approval process similar to the App Store. Or, you know - make the whole situation easier to inspect.