How attackers are jailbreaking LLMs with CTF framing and how to catch them

Sysdig TRT observed multiple threat actors using CTF/CVE-style prompt framing to jailbreak coding assistants and produce working exploit code, which was then deployed against LLM infrastructure (PraisonAI, LiteLLM, FastGPT, Open-WebUI) and other apps (Gotenberg). The technique leaves a distinctive fingerprint—CVE/CTF strings embedded across User-Agent, passwords, API-key aliases, and IAM session names—and was used in active exploitation, credential harvesting, cloud pivot attempts, and remote code execution; defenders are advised to detect and block CVE-templated UA patterns and sanitize fields before feeding events to LLM-assisted SOC tools.