A compromised Fireworks AI access token allowed threat actors to circulate a tampered version of Anthropic’s Claude Code extension, resulting in credential exposure and service interruptions.

Hackers have weaponized a stolen Fireworks AI access token to distribute a malicious version of Anthropic’s Claude Code extension, exposing user credentials and disrupting services.

How the breach unfolded

Threat actors obtained a valid Fireworks AI token, which they used to publish a tampered Claude Code extension on the official marketplace. The compromised extension contained hidden code that harvested API keys and other sensitive data from users who installed it.

Technical details of the malicious extension

The altered extension leveraged the same package name as the legitimate Claude Code tool, making detection difficult. Once installed, it executed a background script that silently sent collected credentials to an external server controlled by the attackers.

  • Steals Anthropic API keys and other tokens
  • Exfiltrates data to attacker-controlled endpoints
  • Can be used to launch further attacks on connected services

Impact on users and services

Affected users reported unauthorized API calls and service interruptions shortly after installing the extension. Organizations relying on Anthropic’s models faced unexpected usage spikes and potential billing surges.

Security teams are advised to revoke compromised tokens, audit recent extension installations, and monitor for anomalous API activity.

For guidance on securing AI integrations, see the Anthropic on Momentum page and explore related resources in the AI category.

"This incident underscores the need for strict verification of third‑party extensions before deployment," said a security analyst familiar with the case.

The Hacker News coverage of stolen Fireworks AI key