Cognition AI Suffers Massive Breach, Proprietary Model Weights and Training Data Leaked

Overview

Severity: CRITICAL | Affected: Cognition AI | Category: breach

Major AI developer Cognition AI has confirmed a significant security breach originating from a compromised third-party data processing library within their MLOps pipeline. Attackers successfully exfiltrated proprietary model weights for their upcoming 'Gen-5' large language model, along with several terabytes of curated training data. The leaked assets, which reportedly include sensitive internal documents and source code, were subsequently posted for sale on a prominent dark web forum. This incident highlights the critical need for robust supply chain security in AI development and underscores the immense value of proprietary models and data as targets for corporate espionage and malicious actors. Cognition AI has initiated an internal investigation and is working with federal law enforcement. The full impact on their competitive standing and future product roadmap is yet to be determined.

References

• https://www.theverge.com/2025/2/15/24012345/cognition-ai-breach-model-weights-leaked
• https://techcrunch.com/2025/02/15/cognition-ai-confirms-theft-of-gen-5-model-and-training-data/