GitBook disclosed on November 3 that an attacker used AI agents to gain unauthorized access to infrastructure credentials, forcing the documentation platform to take authenticated services offline and revoke every customer's API tokens1,3.
The intrusion began at 7 p.m. Central European Time on November 22. GitBook described the incident as "a sophisticated, highly targeted attack using AI agents" and said it detected the breach immediately. The company took its application and all documentation sites requiring authentication offline while it secured its infrastructure.
Exposed data and remediation steps
Potentially exposed information includes API tokens, integration secrets, and questions submitted to GitBook's AI assistant. GitBook said it found no evidence that hosted content was modified. The company replaced its own infrastructure credentials, invalidated customer API tokens, and instructed customers to rotate any third-party API tokens or secrets used to connect external tools to the platform. Customers who use GitBook's Model Context Protocol server were told to re-authenticate.
GitBook has not explained how the attacker deployed AI agents, which AI systems were involved, or how the credentials were initially obtained. Its disclosure does not establish that GitBook's own AI products were the vector for the intrusion.
Downstream fallout
The breach rippled into decentralized finance. PancakeSwap temporarily took its documentation offline as a precaution and redirected visitors to its homepage. PancakeSwap said its documentation was not compromised and its application was unaffected. PancakeSwap later moved its documentation away from GitBook to infrastructure under its own control.
GitBook's status page first reported downtime affecting published documentation sites and the editor on November 3. Publicly published content remained accessible through cached copies for as long as possible during the outage. The company later said its application, API, and published documentation were available again, though some services could remain degraded.
ANALYSIS The incident is notable less for the scale of data exposed than for the attack method GitBook described. AI agents that can carry out sequences of actions and interact with software autonomously represent a different threat profile from conventional credential-stuffing or phishing vectors. GitBook's sparse technical detail, however, leaves the actual role of AI agents in the kill chain unverified. PancakeSwap's decision to migrate its documentation off GitBook entirely rather than simply wait for restoration suggests at least one major customer views the platform risk as structural, not episodic.
The breach follows a period of heightened attention to AI-system security across the industry. OpenAI recently overhauled its bug bounty program and partnered with Cathedral to stress-test its own defenses[2].