Arm Brings Agentic AI to Vulnerability Research: Metis Becomes Open Source

Arm has made available, as an open source project, Metis, a security framework based on agentic artificial intelligence that its product security team uses to identify vulnerabilities in code. The tool is already in use on over 130 internal software projects within the company, with the aim of extending it to all Arm software by the end of 2026.

The official announcement, signed by Mark Hambleton, Senior Vice President of the Software Division, comes only now, even though Metis's code has been public for months: the repository has fifteen releases, the latest in early March. This recent announcement is thus the formal presentation of the tool, accompanied by the first effectiveness data, more than a real debut.

Metis fits into a burgeoning category of tools that use AI to uncover software vulnerabilities before products reach users. This was recently discussed in relation to Anthropic's system capable of identifying vulnerabilities in Windows, Linux, and macOS.

How It Works: Context Instead of Fixed Rules

Unlike traditional static analysis tools, which rely on predefined rules and known pattern searches, Metis is built on a RAG (retrieval-augmented generation) architecture, which combines language models with specific knowledge of the project. The tool builds a knowledge base from source code, build files, and documentation, and reasons on that: it can analyze an entire repository, a single file, or just the changes introduced by a pull request.

This allows, at least according to Arm's claims, to recognize vulnerabilities spanning multiple functions or components, which are the hardest to catch for rule-based tools. Metis can also validate reports produced by other tools, distinguishing likely vulnerabilities from false alarms. In its internal benchmarks, Arm claims up to 10 times more true positives and about 50% fewer false positives compared to leading static analysis tools. While these numbers are from the company and not independently verified, the problem they identify is real: false alarms consume time and erode trust in automation.

In its internal uses, Arm runs Metis on GPT-5.5-Cyber from OpenAI, but the tool is designed to be supplier-independent and works with any endpoint compatible with OpenAI APIs, including locally executable solutions like vLLM and Ollama. Those who do not want to send their code to an external service, or have specific confidentiality needs, can keep everything in-house.

From Software to Hardware, Under Open License

Metis is distributed under Apache 2.0 license and supports a wide range of languages, including C, C++, Python, Rust, Go, and Solidity. The decision to make it available as open source, explains Arm, stems from the idea that security is a challenge for the entire ecosystem, and in this regard, the company reports already having interest from some partners.

However, the more ambitious direction looks beyond software. Arm recently added support for Verilog, the hardware description language, and claims to be working with partners to apply the same method to vulnerability research in chips. For those designing the architectures on which much of mobile processors are based, shifting automatic analysis from code to silicon is a move in line with their craft.

The project is available immediately: code and documentation can be found on Arm's GitHub repository, while details and data are collected in the official announcement.