DARPA AI Forge Cycle 2 Project Ventures

DARPA AI Forge Cycle 2 Portal

All applications for this internal competition must be submitted via InfoReady. You can sign in using your Net-ID and your @stonybrook.edu email address. Please allow sufficient time to familiarize yourself with the portal prior to submitting your pre-application. Pre-application requirements are listed in the announcement below.

Please note: The sponsor will release the Request for Abstracts for this cycle on October 23, 2026. We will update this internal competition accordingly as soon as that information is available.

The Office for Research and Innovation invites applications for Cycle 2 of the DARPA AI Forge program, which aims to support foundational research on emerging AI challenges and national security priorities. Stony Brook University can submit ONE application to this program.

Please see below for details. If you are interested in applying, you must submit a pre-application via the DARPA AI Forge Cycle 2 Portal in InfoReady by Monday, October 26, 2026 at 9 am. Pre-applications for internal review and selection must be uploaded to the portal as a single PDF document.

KEYWORDS: Artificial Intelligence, National Security

WHAT DOES IT FUND: DARPA AI Forge seeks disruptive scientific advances that make frontier AI systems interpretable, controllable, reliable, and secure in high-stakes, adversarial national security missions. DARPA encourages prospective applicants to review the AI Forge Critical AI Challenges for National Security Report - Cycle 2 in preparation for the Cycle 2 solicitation. Riverside Research will publish the Request for Abstracts and related submission guidance on October 23, 2026. The Cycle 2 report places increased emphasis on autonomous agents operating over time with tools, persistent memory, delegated tasks, and changing conditions. Cycle 2 abstracts should emphasize deployed multi-agent systems and cross-model approaches. The program is focused on aligning government, academia, and industry around forward-looking foundational research to accelerate advances in the following nine research challenges across four complementary thrusts in Cycle 2:

Thrust	Research Challenges	Capability Question
1 - AI Interpretability	1.1 Predict Emerging Failure Can we predict emerging divergence, instability, and failure early enough to act? 1.2 Establish Causal Accountability Can we identify root causes of failure and effective corrections?	Can we detect emerging failures early enough to act and establish what caused them and how to correct them?
2- AI Control	2.1 Enforce Authorized Bounds Can we keep agents pursuing authorized missions through permitted means? 2.2 Verify Steerability as Objectives and Conditions Change Can we steer agents as mission objectives change? 2.3 Control Autonomous R&D and Experimentation Can we secure, monitor, and shape impacts from autonomous R&D or scientific experimentation?	Can we maintain effective authority over agents as mission objectives change and automated AI capabilities grow?
3 - Adversarial Robustness	3.1 Preserve Mission Integrity Against Adaptive Adversaries Can we protect mission integrity as adversaries team and adapt? 3.2 Sustain Secure and Available AI Through Compromise and Disruption Can we sustain and restore secure operations through compromise and disruption?	Can we preserve mission capability as adversaries adapt and restore secure operations after compromise or disruption?
4 - AI Assurance* *New thrust focuses on verification, evaluation, and mission confidence. Addresses whether available evidence justifies continued confidence in system.	4.1 Verify Scientific Outputs of AI Systems that Exceed Human Expertise Can we independently verify AI outputs that exceed human expertise or where no ground truth exists? 4.2 Maintain Defensible Mission Assurance Over Time Can we assess defensible mission assurance as conditions change over time?	Can we verify difficult AI results and justify continued trust in a system for a particular mission as conditions change?

Each project abstract should only address one of these challenges. Applicants must demonstrate a clear understanding of the frontier state of the art in the challenge area, show exactly what missing technical capability the proposed research establishes, and explain how their approach complements rather than duplicates frontier-lab research. Applicants must avoid topics that industry is likely to solve in the next 6-12 months.

DARPA will select up to 15 universities for Cycle 2. The final number of full proposals selected will depend on quality of pitches and funding available. DARPA expects to change the challenges each cycle to match the speed of AI.

Please note: Intellectual Property (IP) generated from Project Ventures (awards) is intended to be shared among AI Forge forum participants, preferably under liberal open-source licensing (e.g., the MIT License). By depositing software that is potentially commercially valuable to an open source repository, Stony Brook applicants may be foregoing some personal financial gain that would come to them via the SUNY IP policy from successfully licensing the software/model to an industry partner for a fee, either exclusively or non-exclusively. Furthermore, Stony Brook applicants must distinguish any project proposed under AI FORGE from any active project that they may have with an industry partner.

THIS PROGRAM DOES NOT FUND:

Projects that duplicate frontier-lab research.
Projects that rely on broad claims and lack sufficient detail to show technical reasoning.
Topics that industry is likely to solve in the next 6-12 months.
Projects led by PIs who are not available to attend the in-person Pitch Day in Washington, D.C. on December 11, 2026.
AI-generated abstracts.

APPLICANT REQUIREMENTS: An individual may serve as PI on only one (1) Project Venture abstract submitted in response to this RFA. The PI must commit 20% or more of their time to the Project Venture. Key personnel (PIs, Co-PIs) are limited to U.S. citizens or permanent residents.

AWARD: Not yet disclosed. Cycle 1 proposed budgets were capped at $1,500,000.

COST SHARE: Not required.

SPONSOR DEADLINE: DARPA anticipates releasing cycles every six months, with an accelerated cycle 2 only. Cycle 2 abstracts are due to DARPA November 6, 2026; Pitches and full proposals due December 8, 2026; Live Pitch Day December 11, 2026 in-person in Washington, D.C.

PRE-APPLICATION: A complete pre-application should include the following:

Abstract (no more than 2 pages) that includes:
- Brief description of how the project will address autonomous agents operating over time with tools, persistent memory, delegated tasks, and changing conditions. Projects must emphasize deployed multi-agent systems and cross-model approaches.
  - Strong proposals must identify an unresolved problem, identify why current methods fall short, and present a technically credible approach as opposed to just the applications of AI. Include enough detail to show technical reasoning (do not rely on broad claims).
  - Applicants must demonstrate a clear understanding of the frontier state of the art in the challenge area and show exactly what missing technical capability the proposed research establishes.
  - Explain how your approach complements rather than duplicates frontier-lab research. Applicants must avoid topics that industry is likely to solve in the next 6-12 months.
  - DARPA encourages applicants to explore using subcontractors or specialized partners where needed for frontier relevance and challenge-specific AI expertise, including Defense Industrial Base (DIB) or specialized companies/startups for operational contributions.
Compute resource estimation (no more than 1 page) Proposals must rationalize use of compute. The problem should be at frontier scale, identifying at-scale problems while including rationale which stakeholders will think are realistic and feasible.
List of all project personnel. Note: If approved to apply to DARPA, applicants must submit a Foreign National Justification form to the sponsor for all personnel who are foreign nationals, including permanent residents, detailing the individual's role, proving their technical expertise is critical and irreplaceable, and documenting their U.S. academic standing and residency.
PI's CV/Biosketch

REVIEW CRITERIA: Pre-applications will be reviewed based on how well they address the sponsor’s program objectives. Strong proposals must identify an unresolved problem, identify why current methods fall short, and present a technically credible approach as opposed to just the applications of AI. Applicants must demonstrate a clear understanding of the frontier state of the art in the challenge area and show exactly what missing technical capability the proposed research establishes. DARPA encourages applicants to explore using subcontractors or specialized partners where needed for frontier relevance and challenge-specific AI expertise, including Defense Industrial Base (DIB) or specialized companies/startups for operational contributions.