Threat Brief — 2026-09-04 — LLM API keys in the crosshairs

CISA's ongoing KEV catalog expansion now includes seven actively exploited vulnerabilities, with fresh reporting highlighting that attackers are using the LiteLLM auth bypass (CVE-2026-59822KEV) specifically to steal LLM API keys. This ties the KEV additions to a concrete exploitation pattern: threat actors are targeting AI/ML infrastructure to harvest credentials for large language model APIs. The detail elevates the LiteLLM entry from a generic catalog listing to an active credential-theft campaign.

Top items

• CISA KEV exploitation now linked to LLM API key theft via LiteLLM. A new article adds detail to the ongoing KEV story (first reported 2026-08-17 by CISA): of the seven actively exploited vulnerabilities blacklisted, the BerriAI LiteLLM auth bypass is being used specifically to exfiltrate LLM API keys. This connects the KEV listing to a tangible attack pattern targeting AI infrastructure. CVE-2026-59822KEV is in CISA KEV. (src: Anquanke); (also first reported by CISA)

Themes

AI/ML infrastructure as attack surface. The LiteLLM API-key theft detail reinforces a pattern visible across recent findings — from Claude Code prompt-injection bypasses to OpenAI's Astra model autonomously discovering zero-days to AI-assisted breach investigations. Adversaries are both targeting AI platforms for credential theft and leveraging AI tooling to scale attacks. Defenders should treat LLM API gateways, MCP servers, and AI orchestration layers as high-value assets warranting the same access controls and monitoring as traditional secrets stores.