Skip to navigation

Threat reportMalwareTL-2026-4533

Infostealer-Driven Claude AI Session Hijacking via Stolen Browser Cookies (Vidar, LummaC2, StealC, RedLine, Acreed, AMOS)

highACTIVE

Commodity infostealers running on user endpoints steal browser cookies, session IDs and saved credentials, letting attackers replay authenticated Claude sessions without a password or MFA prompt. Anthropic force-signed-out affected users, removed saved payment methods and refunded unauthorized charges, characterizing the issue as endpoint compromise rather than a Claude infrastructure flaw.

Key facts

IdentifiersTL-2026-4533
Also known asClaude session hijacking, Claude session theft
SeverityHIGH
Affected productsAnthropic Claude (web sessions, accounts and Claude Code OAuth tokens on infected endpoints); Microsoft Windows (infected endpoints: Vidar, LummaC2, StealC, RedLine, Acreed); Apple macOS (infected endpoints: Atomic Stealer)
Weaknesses (CWE)CWE-522, CWE-613
MalwareAMOS, Acreed, Atomic Stealer (AMOS), LummaC2, RedLine, Stealc, Vidar
MITRE ATT&CK techniquesT1005 Data from Local System, T1078 Valid Accounts, T1204.002 Malicious File, T1496.004 Cloud Service Hijacking, T1528 Steal Application Access Token, T1539 Steal Web Session Cookie, T1550.004 Web Session Cookie, T1555.003 Credentials from Web Browsers
MotivationFINANCIAL
Target sectorstechnology, professional services, general consumers
Target regionsGlobal
Detection rules9 (1 shown free)
Sources6 cited sources

Detection coverage

As of 2026-10-27, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-4533 across Splunk SPL, Microsoft KQL and Sigma, covering 7 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Infostealer Browser Cookie/Credential Store Access by Non-Browser Process (TL-2026-4533)

Detects a non-browser process reading two or more browser credential or cookie stores (Chromium Login Data / Cookies / Web Data / Local State, Firefox logins.json / key4.db / cookies.sqlite) within the search window. This is the core collection behaviour of the Vidar, LummaC2, StealC, RedLine and Acreed infostealers behind the Claude session-hijacking campaign (TL-2026-4533): stolen cookies are replayed to take over authenticated Claude sessions, bypassing password and MFA. Requires Windows object-access auditing (SACL) on browser profile directories.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1555.003, T1539, T1005
  • Data sources: windows_security, wineventlog
  • Severity: high
  • Confidence: medium

Splunk SPL

index=wineventlog sourcetype="WinEventLog:Security" EventCode=4663 earliest=-24h
    (ObjectName="*Login Data" OR ObjectName="*Web Data" OR ObjectName="*Local State" OR ObjectName="*Cookies" OR ObjectName="*logins.json" OR ObjectName="*key4.db" OR ObjectName="*cookies.sqlite")
    NOT ProcessName IN ("*chrome.exe","*msedge.exe","*firefox.exe","*brave.exe","*opera.exe","*vivaldi.exe","*MsMpEng.exe")
| stats dc(ObjectName) as store_count values(ObjectName) as stores min(_time) as first_seen max(_time) as last_seen by host, SubjectUserName, ProcessName
| where store_count>=2
| convert ctime(first_seen) ctime(last_seen)

Microsoft KQL

SecurityEvent
| where TimeGenerated > ago(24h)
| where EventID == 4663
| where ObjectName endswith "Login Data" or ObjectName endswith "Web Data" or ObjectName endswith "Local State" or ObjectName endswith "Cookies" or ObjectName endswith "logins.json" or ObjectName endswith "key4.db" or ObjectName endswith "cookies.sqlite"
| where ProcessName !endswith "chrome.exe" and ProcessName !endswith "msedge.exe" and ProcessName !endswith "firefox.exe" and ProcessName !endswith "brave.exe" and ProcessName !endswith "opera.exe" and ProcessName !endswith "vivaldi.exe" and ProcessName !endswith "MsMpEng.exe"
| summarize StoreCount = dcount(ObjectName), Stores = make_set(ObjectName, 10), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by Computer, SubjectUserName, ProcessName
| where StoreCount >= 2

Sigma

title: Non-Browser Process Accessing Browser Credential and Cookie Stores
id: 3b7e1a54-9c2d-4f08-a1e6-5d4c7b2f9a01
status: experimental
description: Detects a process other than a browser opening browser cookie or saved-credential databases, the collection step of Vidar, LummaC2, StealC, RedLine and Acreed infostealers used for Claude session hijacking (TL-2026-4533). Requires object-access auditing on browser profile folders.
references:
    - https://www.esecurityplanet.com/threats/news-claude-session-hijacking-infostealer-malware/
    - https://www.s-rminform.com/cyber-intelligence-briefing/cyber-intelligence-briefing-4-september-2026
    - https://attack.mitre.org/techniques/T1539/
author: Threadlinqs Team
date: 2026-10-27
tags:
    - attack.credential-access
    - attack.t1555.003
    - attack.t1539
    - attack.collection
    - attack.t1005
logsource:
    product: windows
    service: security
detection:
    selection:
        EventID: 4663
        ObjectName|endswith:
            - 'Login Data'
            - 'Web Data'
            - 'Local State'
            - 'Cookies'
            - 'logins.json'
            - 'key4.db'
            - 'cookies.sqlite'
    filter_browsers:
        ProcessName|endswith:
            - 'chrome.exe'
            - 'msedge.exe'
            - 'firefox.exe'
            - 'brave.exe'
            - 'opera.exe'
            - 'vivaldi.exe'
            - 'MsMpEng.exe'
    condition: selection and not filter_browsers
falsepositives:
    - Backup, migration or profile-sync tools reading browser profiles
    - Forensic or DLP agents scanning user profiles
level: high

Known false positives:

  • Backup, profile-migration or browser-sync utilities that read browser profile databases
  • EDR/DLP agents scanning user profile directories

Tuning notes: Enable Audit File System and add a SACL on the Chrome, Edge and Firefox profile directories under each user's AppData. Add approved backup/EDR binaries to the NOT ProcessName list. Raise store_count to 3 to cut noise, or lower it to 1 for high-value users.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Execution of Cracked/Pirated Software Loader or Known Community Stealer Hash (TL-2026-4533) · unlock with TL Creds
  • Chromium Browser Launched With Remote-Debugging Port by Non-Interactive Parent (TL-2026-4533) · unlock with TL Creds
  • AI-Assistant Session Replay and OAuth Grant Abuse in Entra ID (Claude/Anthropic) (TL-2026-4533) · unlock with TL Creds
  • Browser Cookie/Login Database Copied via esentutl or Shell Copy Commands (TL-2026-4533) · unlock with TL Creds
  • macOS Atomic Stealer (AMOS) Keychain Extraction and Password Prompt Activity (TL-2026-4533) · unlock with TL Creds
  • Non-Browser Process Resolving Steam/Telegram Dead-Drop Hosts Used by Stealers (TL-2026-4533) · unlock with TL Creds
  • Infostealer Staging Copy of Browser Credential Databases to Temp (TL-2026-4533) · unlock with TL Creds
  • Connections to Community-Observed Vidar Infrastructure (5tr.org, 37.27.4.20) (TL-2026-4533) · unlock with TL Creds

Indicators of compromise

As of 2026-10-27, Threadlinqs Intelligence tracks 7 indicators of compromise for TL-2026-4533.

7 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Between late August and early September 2026, Anthropic warned Claude users that general-purpose infostealer malware on their own devices was being used to hijack active, authenticated Claude sessions. The S-RM Cyber Intelligence Briefing of 4 September 2026 summarized the issue (sourced from a cybersecuritynews.com article) as malware stealing saved passwords, browser cookies and locally stored credentials, allowing attackers to take over active Claude sessions without passwords or MFA approval. The S-RM briefing does not name malware families; secondary reporting (eSecurityPlanet, TheCyberSecGuru, Paubox, Ecosistema Startup) states Anthropic identified Vidar, Lumma (LummaC2), StealC, RedLine and Acreed on Windows, and Atomic Stealer (AMOS) on macOS, with fewer macOS cases.

Attack chain as reported: a malicious or untrusted download (unofficial apps, cracked or pirated software, software disguised as legitimate, a pirated game cited as an example) installs an infostealer; the stealer harvests browser data including session cookies and authentication tokens; the attacker replays the stolen session on attacker-controlled systems, impersonating the user. Because the cookie represents an already-authenticated state, the login step (and any MFA protecting it) is bypassed. No vulnerability in Anthropic infrastructure is exploited, and Anthropic stated it had no reason to believe the malware was related to Claude, installed through Claude, or caused by anything victims did while using Claude.

Impact reported: attackers consume the victim's paid usage allowance (users noticed usage limits that refilled and then drained while idle; one documented case of a UK AI consultant on a Claude Max 20x plan saw anomalous consumption on 4 August 2026 and unauthorized OAuth tokens created for Claude Code), saved payment methods were affected, and per one source stolen access is resold on underground forums at a fraction of subscription price while the legitimate owner is billed. One analysis reports stolen Claude sessions could reach connected systems through standing OAuth grants (for example corporate Gmail). One source reports web session cookies persist for 28 days of inactivity with hourly refresh, enabling weeks of reuse after theft; this and the claim that Claude lacks native MFA come from a single secondary source and are unverified.

Anthropic's response: forced sign-out of affected accounts (stopping stolen sessions), revocation of stolen tokens, removal of stored payment methods, refunds where unauthorized charges were identified, and email notification of affected users. Signing out does not remove the malware, so victims must clean the endpoint and rotate credentials of linked email accounts. The number of affected accounts has not been disclosed. No network or file IOCs were published in any source, so no BeaconBeagle correlation was possible. The S-RM briefing also covers unrelated items (Rhysida/Berlin, McKesson, AI-automated GitLab ransomware, Sality/TeamPCP, Lenovo/Dropbox, Starlink) that are out of scope for this record.

MITRE ATT&CK techniques by tactic:

Collection

T1005 Data from Local System

Initial Access

T1078 Valid Accounts

Execution

T1204.002 Malicious File

Impact

T1496.004 Cloud Service Hijacking

Credential Access

T1528 Steal Application Access Token; T1539 Steal Web Session Cookie; T1555.003 Credentials from Web Browsers

lateral-movement

T1550.004 Web Session Cookie

Remediation

Immediate actions

  • Stop using the suspected-compromised device for sensitive logins and run a full malware scan (e.g. Microsoft Defender on Windows, macOS security utilities)
  • Secure the email account linked to Claude: set a new password, sign out other devices, enable fresh 2FA
  • Revoke active sessions and API/OAuth tokens, including Google and browser sessions from secondary devices
  • Review financial statements and re-enter payment details only after confirming the machine is clean

Workarounds

  • Clean the endpoint first: signing out of Claude stops stolen sessions but does not remove the malware

Longer-term hardening

  • Deploy EDR with infostealer detection on endpoints
  • Implement session anomaly detection (impossible travel, device fingerprint mismatch, concurrent sessions)
  • Reduce session token lifetimes and establish revocation paths independent of vendor detection
  • Audit standing OAuth grants and delegated access held by AI assistants
  • Block pirated/cracked software and unofficial app downloads via policy

Affected and fixed versions

  • Anthropic — Claude (web sessions, accounts and Claude Code OAuth tokens on infected endpoints)
    Vulnerable versions: Accounts authenticated on endpoints infected with infostealers
    Fixed in: No product patch; Anthropic forced sign-outs and token revocation for affected accounts
  • Microsoft — Windows (infected endpoints: Vidar, LummaC2, StealC, RedLine, Acreed)
    Vulnerable versions: Endpoints running infostealer malware
  • Apple — macOS (infected endpoints: Atomic Stealer)
    Vulnerable versions: Endpoints running Atomic Stealer

Timeline

  • A UK AI consultant with a Claude Max 20x account notices anomalous token consumption during inactivity and unauthorized Claude Code OAuth tokens (later reported to TechCrunch on 2026-09-08)
  • eSecurityPlanet and TheCyberSecGuru report Anthropic's warning that infostealers (Vidar, LummaC2, StealC, RedLine, Acreed, AMOS) are hijacking Claude sessions; Anthropic signs affected users out, removes stored payment methods and refunds unauthorized charges
  • Paubox publishes coverage noting Anthropic emailed affected users and that the number of affected accounts is undisclosed
  • S-RM Cyber Intelligence Briefing summarizes the campaign: stolen passwords, cookies and local credentials let attackers hijack active Claude sessions without passwords or MFA
  • Vijilan analysis published (updated 2026-10-21) describing cookie replay, anomalous-usage detection by Anthropic, and stolen sessions reaching connected systems via OAuth grants
  • Ecosistema Startup reports stolen Claude access is resold on underground forums and describes the documented case reported to TechCrunch

Sources cited

Previous and next in malware

Drafted by the Threadlinqs research pipeline from 6 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats