Anthropic, the company behind Claude AI, has disclosed that threat actors deployed multiple infostealer malware variants to harvest session credentials and unauthorized access to Claude user accounts. The company did not disclose the total number of affected users or provide details about the attack's scope.
The attack relied on commodity infostealers, malware families designed to extract sensitive data from compromised devices. These tools typically target browser cookies, stored passwords, autofill data, and cached credentials. When deployed against Anthropic users, the infostealers successfully captured session tokens and authentication materials needed to access Claude accounts without requiring the actual password.
Session theft attacks represent a particular risk because they bypass traditional authentication controls. Even if a user maintains a strong password and enables multi-factor authentication, a stolen session token can grant attackers immediate access to the account. The attacker can then explore uploaded files, chat history, and any integration data linked to the Claude workspace.
The identity of the threat actor remains unconfirmed. Anthropic has not disclosed whether the infostealer distribution occurred through phishing campaigns, watering hole attacks, or software supply chain compromises. The lack of attribution details complicates threat modeling for organizations relying on Claude for sensitive work.
Infostealer malware campaigns have intensified over the past 18 months. Families like Raccoon, Vidar, and Lumma operate as malware-as-a-service platforms, sold or rented to multiple threat actors on underground forums. These tools extract data from infected systems and upload it to attacker-controlled servers. Session cookies stolen this way sell for between $50 and $500 per batch depending on the target's perceived value.
Anthropic users working with proprietary data face particular exposure. Claude processes conversations for code review, business strategy, legal analysis, and security research. A compromised account grants attackers access to all materials fed into the system. Organizations using Claude for training data or model fine-tuning should audit their session activity and audit logs immediately.
The company has not confirmed whether it invalidated existing sessions across affected accounts or issued mandatory password resets. Anthropic customers should assume their sessions may be compromised and log out of all active Claude sessions within their account settings. Users should also update their device security posture, particularly if they accessed Claude from systems running outdated antivirus software or unpatched operating systems.
The attack underscores a broader challenge for SaaS platforms. Session tokens stored in browser memory or disk remain vulnerable to infostealer families operating on millions of compromised endpoints. No platform can fully defend against compromised client devices. Organizations should implement IP-based access controls and session risk detection where available, restrict Claude API key permissions to minimum necessary scopes, and monitor for unauthorized account activity through audit logs.
Anthropic has not announced changes to its authentication architecture in response to this incident. The company typically uses session-based authentication for its web interface. Additional details about detection methods, timeline, and remediation steps should come from Anthropic in coming days.