This is a SOURCES SOUGHT NOTICE for market research purposes. THIS IS NOT A REQUEST FOR PROPOSALS OR A REQUEST FOR QUOTATIONS.

Artificial Intelligence (Al) is one of the defining technologies of our era. Its emergence, together with its multiplying contexts of use and increasing capabilities, presents enormous opportunities as well as significant present and future harms.

To help encourage innovation while protecting against risks from Al, President Biden signed Executive Order 14110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. The Executive Order tasked the National Institute of Standards and Technology (NIST) with establishing guidelines and best practices for developing and deploying safe, secure, and trustworthy Al systems, including "launching an initiative to create guidance and benchmarks for evaluating and auditing Al capabilities, with a focus on capabilities through which Al could cause harm."1

In particular, the Executive Order focuses on models that might pose a risk to "security, national economic security, national public health or safety," among other things, by "enabling powerful offensive cyber operations through automated vulnerability discovery and exploitation against a wide range of potential targets of cyber attacks."2

The U.S. AI Safety Institute (AISI), housed within the National Institute of Standards and Technology, was created to respond to the priorities assigned to NIST under the Executive Order. AISI is advancing the science, practice, and adoption of AI safety across the spectrum of risks.

As part of this mandate, AISI is conducting pre-deployment testing, evaluation, validation, and verification (TEVV) on frontier models of cyber capabilities and risks.3 In doing so, AISI seeks to ensure that its projects, evaluations, and tools reflect the best available science, and to coordinate closely with a diverse set of Al stakeholders who are developing and conducting evaluations to assess cyber capabilities and risks.

NIST is performing market research to identify potential sources for an anticipated contract to assist in developing evaluations and benchmarks of AI models' relevant cyber capabilities and risks.

POTENTIAL CONTRACT REQUIREMENTS

The following requirements are provided to describe the Government's minimum needs. As the acquisition planning process proceeds, requirements may be modified.

The Contractor must provide or develop resources for various aspects of assessing frontier Al model cyber capabilities and risks. The Contractor would be responsible to conduct one or more of the tasks in list A in order to assess one or more of the capabilities in list B.

Contractor Tasks:

LIST A - Contractors must provide or develop resources for one or more of the following.

1. Developing benchmarks and scoring mechanisms for automated evaluation of Al models' relevant cyber capabilities based on real or realistic offensive cyber tasks or workflows;
2. Developing tasks for automated evaluation of AI models’ relevant cyber capabilities with accompanying data on human baseline performance (e.g., how long the tasks take human experts to complete);
3. Creation of synthetic environments such as cyber ranges that emulate features of realistic networks and support automated grading or scoring for evaluation of AI models’ relevant cyber capabilities;
4. Design and implementation of protocols or methods for evaluating AI models’ relevant cyber capabilities;
5. Development of technical resources or infrastructure for generating tasks and environments and associated scoring mechanisms for evaluating AI models’ relevant cyber capabilities;
6. Development of technical resources or infrastructure for converting existing assets (e.g. publicly available codebases) into evaluations and scoring mechanisms for AI models’ relevant cyber capabilities;
7. Design and implementation of research projects to identify or measure the use of AI models’ relevant cyber capabilities in real-world cyber attacks and offensive cyber operations.

LIST B - Relevant frontier model capabilities to elicit, evaluate, and benchmark include:

1. Capabilities that enable a model to discover vulnerabilities in real or realistic code bases, web resources, or networks;
2. Capabilities that enable a model to develop working exploits for discovered or known vulnerabilities in real or realistic code bases, web resources, or networks;
3. Capabilities that enable a model to automate social engineering workflows, such as the generation of targeted phishing content;
4. Capabilities that enable a model to automate open-source research and/or pre- or post-compromise information gather