NIST is giving AI developers, government agencies and smaller businesses another way to examine how machine learning systems behave under pressure. Its re-released Dioptra tool is built to test how malicious attacks can weaken an AI system, especially attacks that "poison" the data used to train a model.
The point is not to declare any model perfectly safe. The point is to create a shared environment where model performance claims can be tested, compared and tracked when systems face simulated threats.
What Dioptra Is Built To Measure
Dioptra is a modular, open source web-based tool from the National Institute of Standards and Technology, the U.S. Commerce Department agency that develops and tests technology for the U.S. government, companies and the public. It was first released in 2022 and has now been re-released as part of NIST's broader work on AI testing.
The tool is named after the classical astronomical and surveying instrument. In this context, the name fits the purpose: Dioptra is meant to help organizations measure risk in a more systematic way.
NIST says Dioptra can help companies training AI models, as well as the people using those models, assess, analyze and track AI risks. It can also be used to benchmark and research models, giving evaluators a common platform for exposing systems to simulated attacks in a "red-teaming" environment.
One major focus is adversarial behavior. That includes attacks designed to make an AI system perform worse by interfering with the training data or by testing how the model responds under hostile conditions.
Why AI Model Risk Testing Needs Better Tools
AI benchmarks are difficult because many of the most sophisticated systems are not fully visible to outsiders. Their infrastructure, training data and other important details are often kept private by the companies that build them.
That makes it hard for customers, researchers, agencies and the broader public to understand whether a model's performance claims hold up outside a controlled demonstration. Dioptra is designed to give evaluators a practical way to test at least some of those claims against simulated threats.
NIST described the goal directly in its press release: "Testing the effects of adversarial attacks on machine learning models is one of the goals of Dioptra." The agency also said the open source software, available for free download, could help the community conduct evaluations of AI developers' claims about system performance.
That community includes government agencies and small to medium-sized businesses. For those groups, a free, open source AI testing tool can matter because model evaluation can otherwise be difficult to standardize and resource-intensive to perform.
Where Dioptra Fits In AI Safety Work
Dioptra arrived alongside documents from NIST and NIST's recently created AI Safety Institute. Those documents describe ways to mitigate some AI dangers, including abuse of AI to generate nonconsensual pornography.
The tool also follows the U.K. AI Safety Institute's Inspect, a tool set aimed at assessing model capabilities and overall model safety. The U.S. and U.K. have an ongoing partnership to jointly develop advanced AI model testing, which was announced at the U.K.'s AI Safety Summit in Bletchley Park in November of last year.
Dioptra is also tied to President Joe Biden's executive order on AI. That order mandates, among other things, that NIST help with AI system testing. It also establishes standards for AI safety and security, including requirements for companies developing models, such as Apple, to notify the federal government and share results of all safety tests before the models are deployed to the public.
Taken together, these efforts show a clear direction: AI safety work is moving from broad concern toward repeatable evaluation, shared testing environments and more formal scrutiny of developer claims.
What Dioptra Can And Cannot Do
NIST does not claim that Dioptra can remove all risk from AI models. That distinction matters. A tool can reveal weaknesses, but it cannot by itself prove that a system will be safe in every real-world setting.
What NIST does propose is more specific. Dioptra can help show which kinds of attacks may cause an AI system to perform less effectively, and it can quantify the effect on performance.
That makes the tool useful for questions such as:
- How does a model respond when exposed to simulated adversarial attacks?
- Which attack types appear to reduce model performance?
- Can evaluators track and compare those effects in a common testing environment?
- Do developer performance claims hold up under a structured evaluation?
There is also an important limitation. Dioptra only works out of the box on models that can be downloaded and used locally, such as Meta's expanding Llama family. Models behind an API, such as OpenAI's GPT-4o, are not supported for now.
That limitation is significant because many widely used AI systems are accessed through APIs. If evaluators cannot download and run a model locally, Dioptra is not currently the right tool for testing it directly.
The Bigger Evaluation Problem
Dioptra enters a field where evaluation itself remains contested. A report from the Ada Lovelace Institute, a U.K.-based nonprofit research institute that studies AI, found that evaluations alone are not enough to determine the real-world safety of an AI model.
One reason is that current policies can allow AI vendors to choose which evaluations they conduct. If the party being evaluated can pick the tests, the result may not give a complete picture of risk.
That does not make Dioptra irrelevant. It makes its role more precise. Dioptra is a testing platform, not a complete answer to AI safety. Its value is in helping organizations run clearer evaluations, study adversarial attacks and measure performance degradation in a repeatable way.
For AI governance, that kind of shared tooling is a practical step. It gives agencies, companies and smaller organizations a common place to start asking harder questions about model behavior before systems are trusted in public use.