Overview
Severity: CRITICAL | Affected: Cognition AI | Category: breach
Cognition AI, a leading developer of large language models, announced a significant security breach. Attackers gained unauthorized access to their internal data storage, exfiltrating several terabytes of proprietary training data. This data includes curated, high-value datasets used to train their flagship 'Insight-5' model, as well as sensitive internal research documents detailing future model architectures. The breach was discovered after portions of the dataset appeared on a dark web forum. The company stated that customer inference data was not compromised, but the incident raises serious concerns about the security of the AI supply chain and the potential for malicious actors to replicate or poison foundational models. An investigation is underway, with initial reports pointing to a sophisticated social engineering attack targeting a senior data scientist.