# We should just report if an AI model actually survives a serious pentest. Would save everyone time, because that's not happened and won't happen anytime soon and that's all you need to know, because… **By:** Georg Zoeller **Published:** 2026-02-20T23:11:01.270Z **Source:** [LinkedIn](https://www.linkedin.com/posts/georgzoeller_before-launch-perplexity-hired-us-to-test-activity-7430750858285547521-kY9l) --- # Georg Zoeller’s Post [Georg Zoeller](https://sg.linkedin.com/in/georgzoeller?trk=public_post_feed-actor-name) We should just report if an AI model actually survives a serious pentest. Would save everyone time, because that's not happened and won't happen anytime soon and that's all you need to know, because you can't fix this ... you can add classifiers to block the attack signature, but there's an infinite number of permutations for these attacks as there's infinite token space. [Trail of Bits](https://www.linkedin.com/company/trail-of-bits?trk=public_post_reshare_feed-actor-name) Before launch, Perplexity hired us to test the security of Comet, their AI-powered browser assistant. We successfully exfiltrated private emails from Gmail by manipulating the assistant with prompt injection. Starting with an ML-centered threat model, we mapped how Comet moves data between a user's machine and Perplexity's servers. Then we attacked it. Four prompt injection techniques got the agent to comply: fake security mechanisms, spoofed system instructions, user impersonation, and malicious summarization steps. The AI assistant couldn't distinguish between legitimate user requests and attacker-injected instructions embedded in malicious web pages. This review is part of an ongoing security program where we regularly stress-test Comet as it evolves. Contact us to build a similar program: [https://lnkd.in/gtv5JYTb](https://lnkd.in/gtv5JYTb?trk=public_post_reshare-text) Depending on the application domain and where and how it is deployed the go/no-go could be softened but overall I think this is what we need at the moment. Well, I do not know about pentests, but intelligent they are not: [https://www.reuters.com/business/retail-consumer/amazons-cloud-unit-hit-by-least-two-outages-involving-ai-tools-ft-says-2026-02-20](https://www.reuters.com/business/retail-consumer/amazons-cloud-unit-hit-by-least-two-outages-involving-ai-tools-ft-says-2026-02-20?trk=public_post_comment-text)/ But the LLMs can now detect security vulnerabilities in your code https://m.[youtube.com/watch?v=sDpkV\_iEnck&feature=youtu.be](http://youtube.com/watch?v=sDpkV_iEnck&feature=youtu.be&trk=public_post_comment-text) Yay! 🥳 It's '26, and I intend to celebrate every LM fuckup we advised people to be cautious about in '25. Who is with me? Sometimes pain is the only teacher.