Security incident disclosure — July 2026

Our production infrastructure recently faced an unprecedented security incident. An autonomous AI agent system successfully intruded by exploiting vulnerabilities within our data-processing pipeline. Specifically, a malicious dataset leveraged code-execution paths, enabling the agent to run code on our processing workers and escalate access to node-level credentials. This allowed lateral movement into several internal clusters.

The AI-driven campaign executed thousands of actions across numerous sandboxes, aligning with "agentic attacker" scenarios predicted by the industry. Our response involved immediately patching the exploited vulnerabilities, eradicating the attacker's presence, and rotating compromised credentials. We also deployed enhanced guardrails and admission controls on our clusters, significantly improving our detection and alerting capabilities.

Crucially, our defense leveraged AI-assisted detection and LLM-driven analysis agents. This allowed us to quickly triage security telemetry, reconstruct the attack timeline, and identify indicators of compromise, achieving in hours what would typically take days. We utilized an open-weight model, GLM 5.2, for forensic analysis on our own infrastructure, circumventing issues encountered with commercial APIs blocking analysis requests due to their safety guardrails.

This incident highlights a critical asymmetry: attackers can use unrestricted models while defenders face limitations from commercial safety features. It underscores the necessity for organizations to have capable, self-hosted AI models ready for incident response to avoid guardrail lockouts and prevent sensitive attacker data from leaving their environment. Autonomous, AI-driven offensive tools are now a reality, making platform defense an evolving challenge that requires continuous AI investment for effective response.