Microsoft's FORGE Lab uses agentic AI tools to find Rust memory-safety bugs

Microsoft Security's FORGE Lab published research on November 9, 2026, describing how its agentic AI vulnerability-scanning system, codenamed Hyenas, found previously unknown memory-safety bugs that can be triggered from 100% safe Rust code in the Rust standard library. The findings appeared on the Microsoft Security Blog, authored by Yechan Bae and the FORGE Lab.

What FORGE Lab found

• Hyenas, the scanning harness used, is a research instance of the codename MDASH multi-model agentic scanning system.
• Every bug was triggered from 100% safe Rust code in the std and core libraries, with no unsafe code required.
• The team identified three recurring misplaced-trust patterns: trusting code not to panic, trusting user-supplied trait implementations, and trusting OS or system state.
• Specific bugs include a panic-safety issue in MapWindows::clone (rust-lang/rust#156501), a str::replace call with a custom Pattern that can produce non-UTF-8 strings (#156491), a process::exit signal-stack vulnerability present since 2015 (#161018), and a CommandExt::exec environment race condition present since 2021 (#156951).
• Over a three-month audit, FORGE Lab reviewed open-source kernels, runtimes, networking libraries, container technology, and media parsers, submitting 155 internally-validated vulnerability reports across 23 projects.

What it means for people building with AI

The research is a dated, concrete demonstration that agentic AI tooling can find real, previously unknown memory-safety bugs in widely trusted safe Rust code at the standard-library level. That matters for anyone building on Rust, including GitHub Copilot's runtime, Android, and the Linux kernel, and it serves as a proof point for AI-driven security research more broadly.

Reporting and next steps

Multiple findings were reported upstream to the Rust issue tracker, and two remained open at the time of publication. The results will feed into MDASH development for Microsoft Defender and connect to Project Perception, a multi-agent security system. FORGE Lab also notes that the Rust Allocator API, stabilizing in Rust 1.100 on November 12, 2026, requires implementors to never unwind from certain trait methods. The work builds on Microsoft's earlier static-analysis tool Rudra from 2021 and prior discoveries CVE-2020-36323 and CVE-2015-20001.

Sources

• FORGE Lab uses agentic tools to identify memory safety issues in Rust (Microsoft Security Blog, November 9, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.