By Hamza Faisal
A pair of new Windows infostealers is going after the credentials that power AI coding tools, harvesting login tokens for ChatGPT, Claude and Codex that can stay valid long after a victim changes their password. Security firm SpyCloud says it found the stolen material in recovered infection logs, and warns that developers who only reset a password after an infection may still be exposed.
The two malware families, WardenStealer and Aoki, were detailed in a SpyCloud report published on October 30. WardenStealer, a Rust-based infostealer sold as a malware-as-a-service subscription, has appeared in more than 130,000 unique infections recaptured by SpyCloud since it was first observed on August 3. Aoki, a Go-based family first seen in the firm's data on October 4, takes a more organised approach, sorting stolen AI material into a dedicated folder labelled "AISessions".
What the malware takes
Alongside the usual haul of browser passwords, cookies, messaging data and cryptocurrency information, both families collect files belonging to locally installed AI applications. In WardenStealer logs, SpyCloud found Claude Code credentials containing OAuth access and refresh tokens, along with authentication material for connected Model Context Protocol servers and the CLAUDE.md file that holds a user's standing instructions for the assistant. Codex authentication data included an OpenAI API key plus access and refresh tokens, while Claude Desktop yielded configuration and usage information.
Aoki's haul is organised by tool. Its AISessions folder holds separate collections for ChatGPT, Claude Code, Claude Desktop, Codex, GitHub Copilot and Cursor, containing credentials, configuration, token files and cookies for web-backed tools. SpyCloud also identified Aoki samples using temporary files beginning with the prefix %TEMP%\akg_ and communicating with a command-and-control domain, neihite.cc.
Why changing your password is not enough
The central risk is that passwords, session cookies, API keys and OAuth tokens are separate authentication artefacts. Changing one does not necessarily invalidate the others, so a stolen session cookie can let an attacker resume an existing session without entering a password or facing another multi-factor authentication challenge. SpyCloud stresses this is endpoint credential theft, not evidence that ChatGPT, Claude or Codex suffered a breach on their own servers.
The danger is not theoretical. Anthropic said on September 3 that infostealer-captured Claude session keys had been used to mint unauthorised Claude Code OAuth tokens and burn paid usage, and it force-signed out affected users and refunded charges. Gen Digital has independently documented the same shift across multiple infostealer families. And the stakes can be large: a Sysdig investigation estimated in May 2024 that a worst-case abuse of Claude across four AWS regions could generate more than $46,000 in a single day.
A resale market — and how to respond
SpyCloud also found underground sellers advertising AI subscriptions and API credit at steep discounts, including $100 of claimed Codex or Claude API credit for $8.94, with discounts of 40 to 91 percent. Researchers cautioned they could not establish whether every advertised credential came from malware rather than phishing, exposed secrets or fraudulent registrations.
For anyone whose machine may have been infected, SpyCloud's advice is blunt: treat AI-tool credentials as part of the investigation. That means rotating exposed keys and tokens, revoking active sessions rather than relying on a password reset alone, and watching usage for anything unfamiliar. As AI assistants plug deeper into code repositories, cloud deployments and company data, the tokens that let them in are becoming as valuable to criminals as the passwords they used to chase.



