Muse Wonder for Cybersecurity: Strong Detection, Fast Code Scanning

Meta’s newest model shows promising security capabilities. In Averlon’s evaluation, Muse Wonder showed good vulnerability scanning capabilities and strong detection rules generation, while showing a clear weakness in interpreting threat-intel reports.

Overview

Meta released Muse Wonder and Muse Wonder Full on October 12. The family supports multimodal input and a one-million-token context window. Meta reports strong coding and agent performance across benchmarks including Terminal-Bench 4.0, DeepSWE, and SWE-Marathon. Meta’s model card

We evaluated Muse Wonder across three cybersecurity benchmarks to understand how those capabilities translate into security work.

What We Evaluated

Our evaluation covered three distinct capabilities:

• Vulnerability localization (VulnGym, Ji et al., 2026): Given an advisory and repository, identify where attacker-controlled input enters and where the unsafe operation occurs. We report judge-matched accuracy: a task is solved only when both locations match the known vulnerability.
• Vulnerability discovery (RealVuln, Pellew and Raza, 2026): Scan challenging repositories for security flaws. We report F3 Score · judge match, real impact, which emphasizes recall and scores findings against annotated vulnerabilities with real impact.
• Threat intelligence and detection engineering (CTI-REALM, Chakraborty et al., 2026): Complete workflows involving threat-report analysis, MITRE ATT&CK mapping, data-source selection, and detection-query generation. We report overall reward on a 0–1 scale.

The code-scanning evaluations used Codex CLI with a common prompt and harness. The threat-intelligence evaluation followed its established scoring procedure.

Results

Vulnerability localization: Muse Wonder correctly localized 16 of 31 vulnerabilities (51.6%), one fewer than GPT-6 Astra, and ahead of the other five models.

Code scanning and vulnerability discovery: Muse scored 59.7, third of the seven models and 1.4 points below the top score. Muse was also the fastest of the seven models, averaging four minutes per repository, and its run cost about a quarter of GPT-6 Astra’s.

Threat intelligence and detection engineering: Muse scored 0.71, second only to Kimi K3 among the seven models, at about half of Kimi K3’s run cost. Muse’s performance varied across the workflow: detection-query generation was a strength, while interpreting the underlying threat reports was weaker.

Access Muse Through Averlon’s Inference Platform

Averlon’s Inference Platform is purpose-built to understand cybersecurity tasks, select the best-fit open or closed model, and evaluate every result. We’ve now added support for Muse Wonder.

Contact support@averlon.io to get started. We’re offering $50 in credits through October 21, 2026.

Notes

a. Judge match. VulnGym and RealVuln each provide an answer sheet of known vulnerabilities and a script that decides, with fixed matching rules, whether a reported vulnerability matches an entry on it. We found this script-based matching to be erroneous: it often misjudges whether a reported vulnerability is the one recorded on the answer sheet. We therefore use two LLM judges, DeepSeek V4.1 Flash and GLM-5.3 Flash, which each evaluate independently whether an identified vulnerability matches an entry on the answer sheet. A finding counts as a match only when both judges agree. For vulnerability localization (VulnGym), a task counts as solved only when both the entry point and the unsafe operation are judged to match the answer sheet.

b. F3 Score · judge match, real impact. For vulnerability discovery (RealVuln) we keep the benchmark’s F3 score, which weights recall more heavily than precision, but compute it with the two-judge match described above in place of the benchmark’s matching script, and count only the annotated vulnerabilities that have real impact.

References

Chakraborty et al. (2026). CTI-REALM: Benchmark to Evaluate Agent Performance on Security Detection Rule Generation Capabilities. arXiv:2603.13517.

Garg et al. (2026). Unconstrained CTI-REALM with Open-Weights. arXiv:2610.14180.

Ji et al. (2026). VulnGym: Benchmarking Coding Agents for Repository-Level Vulnerability Detection. arXiv:2608.02001.

Pellew and Raza (2026). RealVuln: Benchmarking Rule-Based, General-Purpose LLM, and Security-Specialized Scanners on Real-World Code. arXiv:2604.13764.