LEGION: Benchmarking Language Agents on Sparse-Signal Investigation Under Adversarial Noise
Organizations: Legion Security
Abstract
Autonomous language-model agents are increasingly deployed on discovery-driven tasks where signal is sparse and most retrieved information is irrelevant. Current benchmarks fail to evaluate this regime due to dense signals, cooperative environments, and ground truth typically being only a few tool calls away. We introduce Legion, a benchmark framework coupling a stochastic enterprise-telemetry simulator (Legion-Sim) with a tiered library of 74 expertly-crafted adversarial scenarios (Legion-Bench) and an evaluation protocol that scores investigation efficiency alongside detection accuracy. Scenarios are injected into freshly simulated companies, allowing independent control over confounder density and event volume. We test five agent architectures on two simulated companies and find that while recall is high (0.82-0.88) precision lags behind (0.61--0.72). They fail to find all of the indicators of compromise of the attack or mistake genuine indicators with attack related ones, which is critical in the context of cybersecurity investigation. Failures are consistent across reasoning tiers and persist across confounder density and event volume. We release Legion, the scenario library, performance on common agent architectures, and pipeline for reproducibility and extension.
Figures & tables
| Family | # | T1 | T2 | T3 | Representative templates |
|---|---|---|---|---|---|
| Social engineering / phishing | 11 | 0 | 9 | 2 | Manipulate a user into malicious actions |
| Account takeover / identity | 14 | 3 | 6 | 5 | Operate as a legitimate user via stolen credentials. |
| Endpoint / malware | 13 | 0 | 9 | 4 | Execute attacker code on a host to obtain persistent device control. |
| AD / hybrid identity | 8 | 1 | 6 | 1 | Abuse Kerberos and directory primitives to escalate privilege. |
| Cloud / SaaS | 9 | 0 | 7 | 2 | Misuse cloud IAM, APIs, or admin actions. |
| Data theft / exfil / DLP | 10 | 0 | 4 | 6 | Stage and egress sensitive data through monitored or covert channels. |
| Direct | ReAct | Multi-Agent | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Model | ER | FR | IoC R | IoC P | ER | FR | IoC R | IoC P | ER | FR | IoC R | IoC P |
| GPT-5.6 Terra | 0.83 | 0.89 | 0.79 | 0.75 | 0.82 | 0.90 | 0.79 | 0.72 | 0.75 | 0.85 | 1.00 | 0.67 |
| GPT-5.6 Luna | 0.88 | 0.91 | 0.80 | 0.52 | 0.82 | 0.90 | 0.80 | 0.44 | 0.81 | 0.85 | 0.99 | 0.47 |
| Gemini Pro | 0.78 | 0.82 | 0.75 | 0.71 | 0.76 | 0.84 | 0.82 | 0.65 | 0.64 | 0.37 | 0.94 | 0.85 |
| Claude Opus 4.7 | 0.75 | 0.89 | 0.82 | 0.56 | 0.75 | 0.87 | 0.85 | 0.61 | 0.71 | 0.91 | 0.99 | 0.40 |
| Direct | ReAct | |||||||
|---|---|---|---|---|---|---|---|---|
| Toolset | ER | FR | IoC R | IoC P | ER | FR | IoC R | IoC P |
| Full toolset (baseline) | 0.92 | 0.94 | 0.87 | 0.64 | 0.91 | 0.92 | 0.88 | 0.69 |
| No analysis tools | 0.92 | 0.92 | 0.84 | 0.69 | 0.88 | 0.90 | 0.82 | 0.72 |
| Query-only | 0.85 | 0.86 | 0.81 | 0.66 | 0.84 | 0.89 | 0.83 | 0.70 |
AI Agent Benchmarks
AI agent benchmarks measure how reliably autonomous systems plan, use tools, interact, and complete tasks across controlled and real-world settings.
See all 1,583 papers