RSAC 2026 in March featured over 570 sessions, and AI agents took up a noticeable chunk of the program. Positive Technologies' ML team picked ten talks and checked what had happened to their predictions since — the write-up is on Habr. The centerpiece is Steve Kupchik's talk from Lumia Security, "AIKatz: Attacking AI Desktop Apps for Fun & Profit."
Here's how it works. First, the attacker gets access to the victim's machine — say, via a trojan. Then they scan the memory of a desktop AI app for authentication data: JWTs have a predictable format, so you can search for characteristic fragments. In the researchers' experiments, tokens were extracted from the memory of ChatGPT and Copilot, and a cookie-stored session key from Claude. You don't even need to exfiltrate chat history: with an active session, you can just request it from the service. In the versions studied, ChatGPT and Claude stored conversations as branches, and a new user message could steer the conversation back to the original branch. For more durable influence, the researcher asked the model to save an instruction via its memory feature, then covered his tracks in the chat.
Also in the mix: Microsoft 365 Copilot, from which, per the researchers, individual OneDrive or SharePoint tokens could be extracted. And a separate Lumia finding: DLL hijacking in Claude, where an Intel graphics driver initiated the loading of a missing library; under certain conditions, a planted DLL broke the isolation between users. Intel acknowledged the report as a duplicate of an earlier one, with a public description dating to 2025. Plus slopsquatting: a model keeps hallucinating the same package name, the attacker registers it in a registry and stuffs it with malicious code.
Our read: the story isn't "AI is insecure" — it's how the vendors responded. Microsoft closed the report as informational, arguing that access to process memory within the same user session doesn't cross a security boundary. OpenAI suggested asking Microsoft about Windows architecture. Technically, both are right: if an attacker is already in your session, they can see your files anyway. The question is what comes next — they also get API access as you, complete with chat history, integrations, and paid quotas.
We think this is the first genuinely new class of problems that AI apps have brought onto corporate laptops. Tokens used to live in the browser and the mail client, and years of defenses were built around that. Now the same artifact shows up in a desktop chat app, and neither users nor admins have any habits for dealing with it. Lumia's advice — locking down access to AI app process memory via security descriptors and monitoring access attempts — sounds sensible, but the researcher himself is honest about the caveat: it's no substitute for endpoint hardening, since the attack starts after the machine is already compromised.
The winners: EDR vendors, who now have a fresh excuse to sell you another rule. The losers: IT support teams, who'll have to explain why the corporate chatbot with SharePoint access now requires its own incident review. In our experience, stories like this move the market not when they're written up, but when someone takes a real hit through a stolen session.
The weak spot in the write-up is the lack of version info. The authors admit outright that the public description of the DLL hijacking dates to 2025, so you can't judge current app or driver versions from it. There are no CVE numbers, no patch dates, and no confirmation that the findings reproduce on current builds of ChatGPT, Claude, or Copilot. The vendors closed the reports as informational — meaning no patch is coming, by definition. Plus, the source material itself is a review of other people's talks, not primary research: details should be checked against Lumia Security's publication, not the retelling.
Disclaimer: The token extraction from ChatGPT, Copilot, and Claude memory, the DLL hijacking, and the slopsquatting have not been officially confirmed by any vendor — Microsoft, OpenAI, and Anthropic closed the reports as informational, no CVE numbers or affected build versions have been disclosed, and the source is a single retelling of the talks on Habr, so details should be verified against Lumia Security's publication.
Forecast: within the quarter, at least one major EDR vendor will ship a dedicated rule for suspicious memory reads of ChatGPT, Copilot, and Claude processes — or publicly refuse to and explain why.
Source: habr.com · post in Telegram