Autoregressive inference is a generation method used by large language models where each new token is produced from the tokens already generated. In practice, the model builds text one step at a time, so longer outputs mean more decoding work, more memory traffic, and more latency. Systems often use a KV-cache to avoid repeating all attention calculations, but the process is still sequential.
In cyber security, this matters because many AI attacks and defenses depend on how the model handles each token. Prompt injection can hide instructions that steer later outputs, and agentic workflows can turn a single malicious input into unsafe tool use if the model keeps iterating on untrusted content. Defenders watch token-level behavior, limit tool permissions, and add filtering or monitoring because every extra generation step is another chance for errors, data leakage, or hijacking.