Imagine this: you ask your email AI to summarize your latest messages. What you don't know: a harmless-looking email contains hidden instructions that make the AI send sensitive content from your entire inbox (bank, lawyer, doctor) to an attacker. That is exactly what happened at Superhuman, as Simon Willison reports.

The attack in detail
The security researcher at PromptArmor uncovered a remarkable attack vector:
- The trap: An attacker sends an email with hidden text (prompt injection). The text is invisible to you, but the AI reads it.
- The trigger: You ask Superhuman's AI
Summarize my recent emails. - The manipulation: The AI follows the hidden instructions and reads sensitive emails from your inbox.
- The exfiltration: The AI packs the data into a URL and embeds it as an invisible Markdown image:

- The theft: Your browser loads the "image", but Google Forms stores the submitted data instead.
The insidious part: you notice nothing. You only see the normal summary.
Why did it work?
Three factors came together:
- The CSP allowed
docs.google.com: The Content Security Policy allowed images from this domain. Sounds trustworthy, after all. - Google Forms as a data sink: Forms also accepts data via GET requests, simply through URL parameters.
- AI with full access: The assistant could read all emails and output Markdown.
What developers should learn from this
This case shows a fundamental problem with AI integrations:
Every combination of AI + external content + rich-text rendering is a potential exfiltration risk.
Specifically:
- Content Security Policy (CSP) alone is not enough. Even "trustworthy" domains can unintentionally act as a data sink.
- Markdown/HTML rendering is dangerous when the AI has access to sensitive data.
- Prompt injections are real. If your AI processes external content, you have to expect attackers to try to manipulate it.
What Superhuman got right
Credit where credit is due: according to promptarmor.com, Superhuman treated the problem as a high-priority incident and rolled out a fix quickly. That is the right reaction; many companies would have played the problem down or reacted more slowly.
Conclusion
AI features in products are tempting, but they open up entirely new attack vectors. As developers we have to understand: if we give an AI access to sensitive data and let it process external content at the same time, we are building potential exfiltration paths.
The question is not whether attackers will try prompt injections, but when.
Sources: PromptArmor Research + simonwillison.net