LLMs in Attacker Crosshairs, Warns Threat Intel Firm

Threat intel firm GreyNoise reports that threat actors have been probing misconfigured proxy servers that could provide access to LLM APIs. Between October 2025 and January 2026, GreyNoise’s honeypots captured over 91,000 attack sessions, including assaults linked to two campaigns. The first campaign began in October, exploiting server-side request forgery through ProjectDiscovery’s OAST infrastructure, and spiked over Christmas with an automated signature.

The second campaign started on 28 December and involved 80,469 attack sessions over 11 days, targeting more than 70 LLM model endpoints in search of misconfigurations that could leak access to commercial APIs. Attacks fingerprinted models from OpenAI, Anthropic, Meta, Google, Mistral, Alibaba, and xAI, with test queries kept innocuous to avoid triggering security alerts, according to GreyNoise.

The reports also note exploitation activity tied to CVE-2025-55182 (React2Shell) and CVE-2023-1389, and suggest the campaign may be reconnaissance for a larger exploitation operation.