A security-operations agent matched a frontier model's detection at 650 times less energy, using safeguards instead of model accuracy
AspireAI community · arXiv
Security operations centers have to turn terabytes of alerts, logs and flows into decisions they can trust without paying frontier-model prices for every flow. LanG 2.0 is an agentic system that detects threats, correlates evidence and writes rules, and it asks whether small local models can carry that work when structural safeguards rather than model accuracy provide the guarantees. It adds entity-anchored correlation, human approval for rule changes and a tamper-evident ledger, tested against eight flow datasets from five networks, real XDR alerts and multistage campaigns. Against its predecessor, campaign correlation precision rose from 18.75% to 51.81% and rule acceptance from at most 60.00% to 98.33%; scored against Claude Opus 5.5, it detected attacks on unseen networks at 73.42% F1 with no false positives versus 58.10% F1 with 49.00% false positives, for at least 650 times less energy, with finetuned models statistically matching it for 5.6 to 59 times less energy. It processed 3.33 million records in 6.4 minutes and over 30 million hourly, and its safeguards resisted prompt injection, label poisoning and ledger tampering, though the whole comparison is against the authors' own baselines and flow datasets.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.