Model3h ago
Reasoning Models Vulnerable To Denial-of-Service Attacks
Internal monologues in reasoning LLMs create a new security flaw. Attackers can trigger excessively long reasoning chains to slow systems to a crawl, effectively launching a denial-of-service attack. This vulnerability targets the step-by-step processing used by advanced models for math and coding. Developers must now implement strict limits on internal token generation to maintain stability.