AdaCore Released Open-Source AI Verification Tool
The GNAT Foundry demonstrator allows developers to formally verify AI-generated software changes to ensure functional safety.
Updated on Sept. 25, 2026 in Artificial Intelligence
Live Poll
Do you trust AI-generated code to be used in safety-critical systems?
AdaCore has released the GNAT Foundry: Intersection demonstrator as an open-source project to verify AI-generated code. The tool uses formal proof methods to check software changes against requirements and architecture in high-stakes development.
Why it matters
As AI coding agents accelerate software production, they frequently outpace the ability of engineers to perform manual safety reviews. This tool addresses that gap by providing a deterministic workflow that generates independently checkable evidence for code modifications.
The demonstrator achieves 100% modified condition/decision and statement coverage using SPARK, a language subset designed to prove the absence of run-time errors. Running a complete change request through the system requires 10 GB of free storage and takes approximately 2 hours.
The players
AdaCore
A software company specializing in safety-critical development environments and compilers for the Ada and SPARK programming languages.
The details
The project functions by using deterministic tools to verify that an AI agent has correctly propagated updates through the entire development lifecycle, including requirements, architecture, implementation, and tests. It utilizes SPARK, a language subset that enables mathematical proof of software correctness, to verify that the generated code is free from run-time errors such as buffer overflows. The current release is designed for Ubuntu 24.04 on x86_64 or AArch64 architectures.
Timeline
September 25, 2026: AdaCore released the GNAT Foundry: Intersection demonstrator.
The Tech Race
This project aligns with long-standing efforts in aerospace and industrial safety to enforce strict formal proofs on software artifacts. It marks an early effort to bring these high-assurance standards to the emerging domain of generative AI coding agents.
Developers can download the project via GitHub to experiment with formal verification workflows on Ubuntu 24.04 systems. Note that the demonstrator is not currently certified to any safety standard and is not intended for deployment on public roads.
The takeaway
AdaCore's release highlights the industry push toward verifiable AI output rather than relying on probabilistic generation alone. Developers should monitor for future benchmarks comparing the cost of this formal verification approach against manual peer review processes.
Further reading
For more on the current landscape of automated safety tools, explore our Artificial Intelligence coverage.
Source note: This article includes information reported by eeNews Europe.
Live Poll
Do you trust AI-generated code to be used in safety-critical systems?
