🔍 Read the full analysis: When AI Sessions Are Stolen, Agent Logs May Follow on ThorstenMeyerAI.com
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Check Point Research reported that the Aokigahara malware targets AI-tool session files and developer credentials. Warp said attackers reached a vendor and accessed a limited amount of agent trace and telemetry data; separate researchers said OpenAI fixed a flaw involving hidden reasoning. The incidents involve different systems, and the scope and consequences of some remain unclear.
Security researchers and AI companies reported three separate incidents in late October involving stolen AI-tool sessions, exposed agent traces and a flaw that could reveal hidden model reasoning. The developments show that AI work can leave sensitive information in several places—from files on a user’s computer to records held by service providers—though the incidents have different causes and their full scope is not known.
On October 23, Check Point Research published an analysis of Aokigahara, an infostealer with a module it calls “aisessions.” The researchers said the malware searches for authentication, session and configuration files associated with AI tools including Claude, ChatGPT Desktop, Codex, Cursor, Windsurf, GitHub Copilot and Gemini CLI. It also targets developer credentials such as Git files, SSH keys and cloud-service credentials. The report does not establish how the malware is distributed or how many devices have been affected.
Warp said attackers compromised a vendor in its infrastructure and accessed a limited amount of agent trace and telemetry data. In a post on October 26, Warp CEO Zach Lloyd put the upper bound at roughly 1% of users. That is an estimate of possible user exposure, not a confirmed count of affected accounts. The public account cited in the source material is Lloyd’s post; further details about the vendor and the data involved were not provided there.
Also on October 23, according to security researchers at Xint, OpenAI fixed a flaw involving its conversation-compaction feature that could be steered to reproduce hidden reasoning in readable text. Xint said OpenAI paid a $6,000 bug bounty. The source material says there was no public OpenAI statement about this specific issue, so the account of the flaw and fix is attributed to the researchers.
Infographic 01 · One infostealer, many keys
Your AI tool’s login files are now on the list.
Aokigahara steals the classics, plus a dedicated module for AI tools and a second one for developer and cloud keys. The analyzed paths are Windows paths.
The AI module: “aisessions”
Login, session and settings files the malware looks for. Not every file holds a secret.
Claude Code + Claude DesktopCodexChatGPT DesktopCursorWindsurfKiroCodeiumGitHub CopilotVS CodeGemini CLIJetBrains AI AssistantAmazon QAWS CLI + SSO
It also queries the Windows Credential Manager.
The developer toolbox: “ide”
What lets an attacker move from one laptop to code and cloud.
- Git credential files
- Files in the SSH folder that may contain private keys
- AWS, Azure and Google Cloud credential and config files
- GitHub Copilot storage
The classics
- Browser logins and cookies
- Crypto wallets
- Password manager files
- VPN and FTP settings
- Screenshot and clipboard
Why AI sessions
- Access “without entering a password”
- Private conversations, uploaded code
- Paid usage and account resale
- Extortion with business data (possible)
Not known
- How it reaches victims
- How many machines are infected
- How operators make money (“not confirmed”)
A stolen session is a festival wristband: nobody checks your ticket again. That is why two-factor login does not help once it is gone.Thorsten Meyer · my reading
Source: Check Point Research, 23 Oct 2026 (malware analysis; counts are mine). Naming a tool as a target does not mean its maker was breached.
Powered by Thorsten Meyer AI
AI Credentials and Traces at Risk
The incidents matter because AI assistants and coding agents can interact with private conversations, source code and connected services. A stolen session token may let someone access an account without repeating its normal sign-in steps, while agent traces can record what a tool read, executed or received as output. The information exposed in any one case depends on the files or logs involved and the access granted to the tool.
Check Point said stolen AI sessions could be used to access private conversations or uploaded code, misuse paid accounts or support extortion attempts. Those are potential uses identified by the researchers, not confirmed outcomes from the Aokigahara samples. Warp’s reported vendor incident points to a separate exposure route: data may be held by an infrastructure provider even when a user’s own device is not compromised.
The reported OpenAI issue adds a third category: information associated with a model’s hidden reasoning. A study cited in the source material, published in August, reported decoding 315,320 encrypted reasoning blocks found in public repositories; its authors identified 367 pieces of personal data and 182 credentials. Those figures concern the paper’s dataset, not the October OpenAI flaw, and do not establish that the flaw exposed the same material.
Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB-C MagSafe Charging Case

Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
As an affiliate, we earn on qualifying purchases.
Three Different Exposure Routes
The reports describe distinct events, not evidence of a single coordinated attack. Aokigahara is malware that searches local files on a computer. Warp’s incident involved an attacker reaching a vendor and accessing some company data. The OpenAI issue, as described by Xint, involved a feature that could expose hidden reasoning through a software flaw. Their shared concern is that AI-related data can persist beyond a live conversation—in session files, logs, telemetry or model-related records.
Check Point’s analysis says Aokigahara also collects browser logins and cookies, cryptocurrency wallet data, password-manager files, VPN and FTP settings, screenshots and clipboard contents. Its AI-targeting module appears alongside those conventional targets. The researchers noted that the malware’s behavior changed across versions and said the files it targets do not necessarily contain usable secrets. Their report also says it does not confirm how the operators monetize stolen data.
The source account interprets the listed file paths as indicating a Windows-focused campaign. That is an interpretation based on the paths, rather than a separate statement by Check Point about the full range of affected operating systems. The researchers supplied sample fingerprints for defenders, but the material cited here gives no infection total.
““should be treated as sensitive assets alongside browser sessions””
— Check Point Research
Scope and Impact Still Unclear
Several key details remain unconfirmed in the cited material. Check Point did not report how Aokigahara reaches victims, how many devices are infected or whether its operators have used stolen AI sessions for any of the suggested purposes. A file targeted by the malware may not contain an active token or other usable credential.
For Warp, the roughly 1% figure is an upper bound given by the CEO, not a confirmed number of affected users. The public account cited here does not specify which vendor was breached, exactly what each accessed trace contained, or whether the data was used beyond being accessed. For the OpenAI issue, Xint reported a fix and bounty, but the source material does not include an OpenAI statement confirming the details or describing potential exposure.
The August study’s counts relate to reasoning blocks published in public repositories. They should not be treated as evidence that the October flaw affected those blocks or that the same types of information were exposed in OpenAI’s systems.
Containment and Disclosure Steps
Users and organizations can review active sessions and revoke those they do not recognize, rotate exposed API and cloud credentials, and keep secrets out of prompts and agent-accessible files where possible. Security teams can check devices for malware and review endpoint alerts, while developers should treat AI session files and agent traces as sensitive data rather than ordinary application logs.
Organizations that use AI services can ask providers and infrastructure vendors what traces and telemetry they retain, who can access them, and how long they are stored. Shorter retention and managed devices may limit the amount of information available after a compromise, but do not by themselves establish that an account or system is safe.
Further clarity depends on any additional disclosures from Check Point, Warp and OpenAI, as well as details from Xint about the reported flaw. The material available so far does not say whether there were confirmed downstream account compromises, misuse of Warp data or exposure of customer reasoning through the OpenAI issue.
Key Questions
What does the Aokigahara malware target?
Check Point Research said it searches for AI-tool session and configuration files, as well as browser data, password-manager files and developer credentials such as SSH, Git and cloud-service credentials.
How many Warp users were affected?
Warp CEO Zach Lloyd gave an upper bound of roughly 1% of users in describing the possible exposure. The cited account does not provide a confirmed affected-user count.
Did OpenAI confirm the reasoning-data flaw?
The source material says Xint researchers reported that OpenAI fixed the issue on October 23 and paid a $6,000 bounty. It also says there was no public OpenAI statement about this specific bug.
Does two-factor authentication stop stolen sessions?
Not necessarily. A stolen active session may allow access without repeating the login process where the second factor is checked. Revoking sessions and rotating potentially exposed credentials can address different parts of that risk.
Were the three incidents part of one attack?
The reports describe three separate developments: malware targeting local files, a vendor compromise reported by Warp and a software flaw described by Xint. The cited material provides no evidence that they were coordinated.
Source: ThorstenMeyerAI.com
