MASTA multi-agent system models protocols for Tamarin
Original title: MASTA: A Feedback-Scheduled Multi-Agent System for End-to-End Tamarin Modeling and Analysis
On Qwen 3.5-plus, MASTA beats OpenCode by up to 0.27 and ranks first versus Codex and Claude Code.
MASTA splits Tamarin protocol modeling among an Architect, an Auditor, and an Arbiter that only schedules work from verifier verdicts. A benchmark pairs 10 protocols with 139 mutation variants and five correctness and completeness metrics. With Qwen 3.5-plus, MASTA improves over OpenCode on the same backbone by up to 0.27 and ranks first on correctness and completeness averages, ahead of Codex (GPT-5.5) and Claude Code (Opus 4.7). Ablations on five protocols show that removing the Arbiter drops property and verification scores to zero.
Unconfirmed: Gains are on the authors’ 10-protocol benchmark, not independent protocol audits.
- The benchmark has 10 protocols and 139 mutation variants.
- Metrics M1–M3 score correctness and M4–M5 score completeness.
- MASTA uses a weaker backbone yet outranks Codex and Claude Code on averages.
- Removing the Arbiter collapses property and verification scores to zero.
- Products
- TamarinOpenCodeCodexClaude Code
- AI models
- Qwen 3.5-plusGPT-5.5Opus 4.7
From the original article
1 min read at arxiv.orgTamarin-based formal verification of security protocols requires expert-authored multiset rewriting rules and temporal-logic lemmas.
By Yongkang Xiao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.