Cut LLM tokens 94% on supply-chain malware scanning by pre-filtering: 93.1% F1 across five languages, 564 new malicious repos found
arXiv 2608.03232·low signal
MalTotal detects malicious code poisoning in package ecosystems in a language-agnostic way, averaging 93.1% F1 across five languages while consuming 94.0% fewer LLM tokens than comparable approaches, and it surfaced 564 previously unknown malicious repositories in the wild. The reusable pattern is the cost architecture: a cheap deterministic filter in front of the model, with LLM judgment spent only on survivors. Relevant to anyone running a recurring scan where per-token cost, not accuracy, is the binding constraint.