What is Arm's Metis agentic AI security framework, and how does it use LLMs and RAG to perform deep security code reviews across 130+ internal projects, achieve a ~95% true positive rate in detecting logic errors and design flaws in C, C++, Python, Rust, and TypeScript, and why h

Scope: 130+ Internal Projects and Language Support

Arm is using Metis internally across more than 130 software projects . The initial launch focused on C and C++ , but the plugin system allows adding support for Python, Rust, TypeScript, Solidity, and other languages . The framework's architecture is language-agnostic at the plugin level.

Performance Claims

Arm's internal benchmarks show that Metis delivers up to 10x higher true positive rates and approximately 50% fewer false positives compared to leading static analysis tools . However, I could not find a source that specifically cites a "~95% true positive rate" as a headline number. Arm's published materials emphasize relative improvements over existing tools (10x better TP rate, ~50% fewer FPs) rather than an absolute 95% figure. The ~95% number may come from a specific internal project subset, a third-party analysis, or a more recent update — but it is not confirmed in the primary sources reviewed here.

Why Arm Open-Sourced It Under Apache 2.0

Arm has stated that open-sourcing Metis is part of a broader strategy to help the entire software ecosystem improve security at scale . The Apache 2.0 license is a permissive, business-friendly open-source license that allows anyone — including other companies, open-source projects, and individual developers — to use, modify, and integrate the tool without restrictive copyleft obligations . The rationale includes:

• Rising AI-driven vulnerability discovery — AI models are dramatically accelerating vulnerability scanning, but they also generate a flood of low-quality, high-noise reports that overwhelm maintainers . Metis aims to solve the quality problem by using RAG to reduce false positives while catching deeper flaws.
• Ecosystem leverage — By making the tool freely available, Arm benefits from community contributions, broader testing across diverse codebases, and improved security in the software supply chain that Arm's hardware depends on .
• Industry collaboration — Open-sourcing under a standard permissive license aligns with Arm's broader open-source strategy and invites adoption by other product security teams, bug bounty programs, and CI/CD pipelines.

Key Caveat

The evidence confirms Metis's architecture (LLM + RAG + agentic review loops), its deployment across 130+ internal projects, its support for multiple languages via plugins, its open-source Apache 2.0 release, and its benchmarked improvements (10x higher TP rate, ~50% fewer FPs). The specific "~95% true positive rate" figure was not found in the primary Arm sources examined here — it may reflect a specific subset of results, a community measurement, or an extrapolation. If that exact number is important to you, I would recommend checking the Metis GitHub README or the latest Arm newsroom post for the most precise internal metrics.