LanG 2.0 Security Agent Benchmarked Against Claude Opus 5.5

A team of four researchers, led by Anes Abdennebi, posted a paper describing LanG 2.0, an agentic system built for security operations centres that detects threats, correlates evidence across alerts and writes detection rules with a human approving each rule change before it takes effect. The system also keeps a tamper-evident ledger of its own decisions. The authors’ central question was whether a small, locally run model can replace a frontier model for most of a SOC’s daily work when the safety guarantee comes from the system’s structure, not from the model being more accurate than everything else on the market.

The paper tested LanG 2.0 against its own prior version and against Claude Opus 5.5 across eight flow datasets drawn from five networks, alongside real extended detection and response (XDR) alerts and simulated multistage attack campaigns. Against its predecessor, the system’s precision at correlating a multistage campaign rose sharply, and the share of its proposed rules that a human reviewer accepted rose to near total. Against Claude Opus 5.5, the comparison was narrower and more specific: scoring network flows against what is normal for that network, LanG 2.0 found attacks on networks it had never seen at a higher F1 score than Claude Opus 5.5 managed, and did so with no false positives, where Claude Opus 5.5 raised a false alarm on roughly half its flagged cases. The authors also report that finetuned local models matched LanG 2.0 statistically on alert classification and rule writing while using a fraction of the energy, and that the full pipeline processed over three million records in under seven minutes, scaling past thirty million an hour.

The numbers

The figures above come from the paper’s own benchmark tables: the jump in campaign correlation precision, the near-complete rule acceptance rate, the detection F1 score on unseen networks, and the false positive rate recorded for Claude Opus 5.5 on the same task. Taken together they describe a system the authors position as a cheaper substitute for frontier-model triage, not a claim that the frontier model performs worse in general. For a SOC lead reading this, the real decision is not which model is smarter but which error profile a shift queue can absorb: more missed attacks with zero noise, or fewer missed attacks buried inside a stream of false alarms that cost human attention regardless of which model is billed per token.