In a striking demonstration of cybersecurity risks, AI startup Cathedral penetrated Anthropic’s network in less than 12 hours using AI agents, exposing employee information and internal communications through flaws in third-party software integrations.
- AI agents exploited third-party software vulnerabilities to breach Anthropic’s network.
- Sensitive employee and partner information was accessed, but not AI model code or weights.
- Anthropic detected the intrusion, cut access, and is investigating possible bug bounty rule violations.
What happened
Cathedral, an AI startup founded by ex-US government officials, utilized AI-driven agents to break into Anthropic’s network in less than 12 hours after manual attempts failed. The entry point was vulnerabilities in third-party business applications integrated within Anthropic’s infrastructure.
Once inside, Cathedral’s agents accessed a limited range of systems, including sales, human resources, and IT databases. Information obtained included employee meeting transcripts, salary details, job offers, as well as internal Slack messages. Notably, Anthropic’s AI model source code and weights were not compromised.
Why it matters
This breach underscores the growing cybersecurity challenge of securing complex enterprises against attacks that exploit third-party software weaknesses. Anthropic, a leading AI developer valued at $1.4 billion and preparing for a public offering, faces increased scrutiny and reputational risk due to this exposure of sensitive internal data.
The incident also highlights the evolving role of AI in offensive cybersecurity techniques, as Cathedral deployed autonomous AI agents to rapidly uncover and exploit flaws. The involvement of a government veteran-founded company adds weight to concerns about nation-state level cyber threats leveraging similar approaches.
What to watch next
Anthropic’s response to the breach, including how it enforces bug bounty policies and manages vendor risks, will be critical to watch. The company has indicated that Cathedral may have violated the bug bounty program’s rules by not fully disclosing the scope of its activity, which could impact any rewards or legal actions.
More broadly, companies integrating AI technologies and third-party software must continuously improve security maturity, with particular attention to vendor risk management. Observers will also look to see if this incident influences Anthropic’s IPO timeline or investor confidence.