Token torching: new AI security threat identified

Security researchers have identified a new threat to AI systems called token torching, which attacks AI models by consuming excessive computational tokens to degrade performance or cause denial of service. Tokens are the units of text that AI systems process; token torching exploits this fundamental aspect of how large language models and similar systems operate by generating input designed to maximize token consumption. The threat bypasses many traditional security approaches because it does not require breaking into a system or compromising data; instead, it leverages the normal operation of the AI inference process—where the system must respond to all queries—against itself. As organizations scale their deployment of AI in production environments, this emerging threat represents a gap in current security and monitoring infrastructure. Companies running large language models face a dual risk: the immediate impact of service degradation or unavailability, and the secondary cost impact of paying for massive computational resources consumed by an attack. Understanding and mitigating token torching is becoming critical for organizations that operate AI systems at scale, requiring new monitoring practices and resource-allocation strategies that can distinguish between legitimate high-compute workloads and malicious token consumption.

Token torching emerges as a novel category of AI security threat at a time when organizations are rapidly deploying large language models and other AI systems into production. Unlike attacks targeting model weights or data, token torching exploits the inference phase—the computational step where an AI produces answers—by forcing excessive token consumption. This attack pattern exploits a structural vulnerability in how modern AI systems are architected: they must process and respond to all input, and an attacker can craft prompts or interactions designed to maximize token usage. The threat is significant because it sits at the intersection of availability and cost; an organization may maintain uptime while incurring unsustainable operational expenses or experiencing service degradation under load. Current security tools, typically designed to detect malware or unauthorized access, do not necessarily flag resource-exhaustion patterns as malicious.