DARPA and NSF Just Launched AI Forge — a Jointly Governed University Forum Funding Interpretability, Control, and Adversarial Robustness Research. The RFI Closes June 22, Project Ventures Pay \$750K–\$3M, and the CAISI Partnership Reshapes the AI Safety Funding Map.

On June 1, 2026, DARPA and the National Science Foundation jointly announced AI Forge — a new university-led research forum focused on AI interpretability, AI control, and adversarial robustness for national security applications. The launch announcement was accompanied by a Critical AI Challenges for National Security report that lays out 15 specific research challenges across the three thrust areas, and by a Request for Information posted on sam.gov asking U.S. universities to describe their AI research expertise, computing infrastructure, national security partnerships, and technology transition experience. The RFI closes June 22, 2026, at 5:00 PM Eastern. Project Ventures grants are expected to range from approximately $750,000 to $3 million, with project durations up to one year and multiple awards expected annually. The forum itself will be administered by a nonprofit and is targeted to launch in summer 2026.

The structural choices in AI Forge — a jointly governed forum, nonprofit administration, mandatory open-source intellectual property sharing, CAISI involvement, university-only eligibility — depart materially from the standard DARPA BAA model and from the standard NSF directorate model. For university research administrators and AI faculty considering whether to respond to the RFI, the operative question is not whether the forum will fund interesting research. It will. The question is whether the institutional commitments AI Forge is asking for — particularly the open-source IP expectation and the national-security positioning — fit the institution's broader AI strategy. This piece works through the program structure, the three research thrusts, the 15 priority challenges, what the RFI is actually asking, and how strong responses should be organized.

The Granted News brief on the launch is at Granted News.

The Forum Model: Why AI Forge Is Not a Standard BAA

DARPA's standard funding mechanism is the Broad Agency Announcement — a topic drop with a defined window, defined evaluation criteria, defined award sizes, and direct DARPA-to-performer contracting. NSF's standard mechanism is the program solicitation, managed by directorate program officers, evaluated by panels, awarded as grants or cooperative agreements directly to institutions. AI Forge does not look like either.

The forum is jointly governed by DARPA and NSF, with CAISI at NIST as a third partner, and administered by a nonprofit that has not yet been publicly named at the time of the RFI. The forum's stated purpose is to "fund, guide, and manage" university-led research projects — language that suggests an active management role for the forum body, not the arms-length project oversight that characterizes most DARPA and NSF awards. Project Venture intellectual property is "expected to be shared among forum participants, preferably through open-source licensing approaches" — a stronger pre-commitment than typical DARPA IP terms and than NSF's standard expectations.

The forum model has direct precedents in the DARPA portfolio. The Information Innovation Office has historically used consortium-style management for several large programs, and the Defense Sciences Office has tested federated research models. The NSF AI Institutes operate under cooperative-agreement structures that vest substantial program direction in the awarded institution. AI Forge appears to draw from both lineages while adding a national security overlay through CAISI and the Department of War / Intelligence Community engagement that the announcement materials reference.

The implication for responding universities is that AI Forge is not a one-time grant relationship. It is an ongoing forum membership. Institutions that respond to the RFI are signaling willingness to participate in a multi-year structure that involves frontier AI companies, defense and intelligence agencies, and other universities, with shared IP and joint research direction. The strategic decision is not whether the institution can do the research — it is whether the institution wants the relationship.

Three Thrusts: Interpretability, Control, Adversarial Robustness

The three research thrusts are the foundation of the program. Each maps to a long-standing AI safety research agenda that has matured substantially over the past three years, and each has a national security framing that distinguishes AI Forge from the comparable basic-research AI safety funding lines at the NSF AI Institutes (deep dive on NSF AI funding here) and the academic AI safety labs.

AI Interpretability. The forum frames interpretability as making "the behavior, decisions, and impacts of AI systems understandable to humans," with the explicit goal of moving beyond routine-setting explanations toward operational interpretability. This is a more ambitious framing than the post-hoc explainability research that has dominated the field through 2024 — it asks researchers to deliver interpretability that holds up in adversarial, high-stakes, and out-of-distribution conditions. The mechanistic interpretability research line, the circuit-level analysis tradition, the activation-engineering and steering-vector lines, and the formal verification approaches to neural network behavior all fit the thrust. So do the chain-of-thought faithfulness lines and the model-organism-of-misalignment approaches that have come out of the AI safety labs over the past two years.

AI Control. The forum frames control as developing tools that "provide strong, verifiable evidence of bounded, auditable, and reliable model behavior" and that establish foundations for maintaining human oversight of advanced systems. The framing pulls directly from the AI control research agenda that has emerged from Redwood Research and from several academic groups working on what they call "scalable oversight" — protocols and architectures that work even when the underlying model is more capable than the overseer. Researchers working on debate protocols, recursive reward modeling, AI-assisted oversight, and constitutional approaches all fit the thrust. Researchers working on formal verification of bounded behavior, hardware-rooted attestation of model identity, and runtime monitoring approaches fit as well.

Adversarial Robustness. The third thrust is the most established research line of the three — adversarial robustness research has had continuous federal funding for over a decade. The forum framing emphasizes resilience "by design so that it maintains its integrity even when under deliberate attack," which signals an interest in approaches that go beyond empirical robustness benchmarks. Certified robustness research, formal robustness verification, robust training approaches, and the security-oriented red-teaming traditions all fit. The national security framing pulls in physical-world adversarial attacks, multi-modal attacks, and attacks on agentic systems that go beyond the standard adversarial-example benchmarks.

The 15 priority research challenges are distributed across these three thrusts and are scheduled to be revisited every six months. The six-month revisit cadence is unusually fast for a multi-year program and signals that the forum body — DARPA, NSF, CAISI, the nonprofit administrator, and the participating institutions — will be actively re-prioritizing as the field moves.

What the RFI Is Actually Asking

The RFI on sam.gov is not a proposal. It is a capabilities statement that will populate "a repository of U.S. universities interested in accelerating next-generation AI research." Submissions are limited to one authorized response per institution, including the military service academies. The four content areas the RFI requests are AI research expertise, computing infrastructure, national security partnerships, and technology transition experience.

The structure of those four asks is informative. Research expertise alone does not satisfy the RFI. AI Forge wants institutions that can actually run the research — meaning sufficient compute to train and evaluate frontier-adjacent models — and that can carry the research through to operational impact through existing relationships with defense, intelligence, and AI standards bodies. The technology transition experience ask is the most consequential and is the easiest place for otherwise strong AI departments to under-respond. Universities that have published heavily on interpretability or robustness but have never moved research into deployed government systems should be explicit about whether they have the institutional infrastructure to do so under the forum model.

The strongest responses will treat the RFI as a strategic-positioning exercise, not as a recitation of lab capabilities. The four content areas should be answered in a way that lets the forum body see how the institution would contribute to the multi-year forum structure — what specific challenges from the 15-challenge list the institution can lead on, what computing infrastructure can be brought to bear, what existing CAISI, NIST, DARPA, or IC relationships can be activated, and what concrete examples of prior research transition exist.

Project Ventures: The Funding Vehicle

The Project Ventures grants are the financial output of the forum. The $750K–$3M range covers a typical one-year PI-led project at a research university, with the upper end accommodating multi-PI projects, postdocs, graduate students, computing budget, and modest equipment. "Multiple awards expected annually" suggests the forum is sizing for a steady-state of roughly 20–60 active projects, depending on the actual award distribution and program budget — which has not been publicly disclosed.

The one-year project duration is short by AI safety research standards. It signals that AI Forge is funding focused, deliverable-oriented work rather than open-ended exploratory research. Researchers should plan proposals around what can be substantively delivered in twelve months, with clear demonstrations or evaluations rather than open-ended methodology development. The shared-IP and open-source expectations should be priced into the project design from the start.

The funding sequence is unusual: the RFI in June establishes the repository, the forum launches in summer 2026, and Project Ventures awards will presumably flow from the forum body to selected universities through whatever mechanism the nonprofit administrator establishes. The mechanism is likely to look more like a cooperative agreement than a traditional grant or contract, given the joint governance and the active management role the forum is taking.

The CAISI Layer: Why NIST Involvement Matters

The Center for AI Standards and Innovation at NIST is the third partner in AI Forge. CAISI was established under the 2025 reorganization of NIST's AI work — the rebrand and refocus of what had previously been the AI Safety Institute. CAISI's mandate is to develop voluntary AI standards and to coordinate AI evaluation infrastructure across the federal government. Its inclusion as a co-leader of AI Forge — rather than as a downstream beneficiary — is a deliberate choice.

The implication for AI Forge research is that the interpretability, control, and adversarial robustness work the forum funds is expected to feed directly into the AI evaluation and standards work CAISI is building. Researchers responding to the RFI should think about whether their work outputs will be in forms that can plug into NIST evaluation harnesses, benchmark suites, and standards documents — not just academic publications. The IP-sharing and open-source expectations align with that direction: the research outputs need to be usable by the broader CAISI evaluation infrastructure, by frontier AI companies' internal evaluation teams, and by the defense and intelligence communities that AI Forge serves.

This is a meaningful departure from the academic-publishing-and-open-source-code rhythm that has characterized most university AI safety research. Researchers who have built their reputations on conference publications and code releases will not need to abandon that work, but they will need to package outputs in forms that the federal evaluation infrastructure can ingest.

Where AI Forge Fits in the 2026 AI Funding Landscape

AI Forge launches into a federal AI funding landscape that has been substantially reshaped over the past twelve months. The OMB Uniform Grant Regulation rewrite (deep dive here) is restructuring how all discretionary federal grants are reviewed and managed. NSF's SBIR/STTR program just reopened with $250 million and a new Strategic Breakthrough Awards tier (deep dive here). DARPA's Defense Sciences Office is running an active SBIR XL pre-release cycle with the July 22 deadline. The NSF AI Institutes are mid-cycle.

AI Forge does not directly compete with any of those. It is structurally different — university-only, forum-administered, open-source IP, national security focus, joint DARPA–NSF–CAISI governance. For universities deciding where to put their AI safety research bets, AI Forge is best understood as the federal-government-aligned channel for the kind of safety research that has historically been funded by Open Philanthropy, Anthropic, OpenAI's superalignment-era grants, and the Frontier Model Forum. The difference is that AI Forge brings with it the national security framing, the CAISI standards-track output expectation, and the federal procurement infrastructure that the private-philanthropy funding has not.

What to Do Before June 22

The institutional checklist for the next nine days is short. The senior research administrator and the AI department leadership need to agree on whether to respond. The four content areas of the RFI — research expertise, computing infrastructure, national security partnerships, technology transition experience — need to be drafted with input from the actual labs that would lead the work, not from a generic research-office capabilities template. The single authorized response should be cleared through whatever institutional review chain handles defense-related submissions, and submitted on sam.gov by 5:00 PM Eastern on June 22.

Institutions that do not respond to the RFI will not be foreclosed from the forum — capabilities repositories typically remain open for additions after the initial response window — but they will not be in the initial cohort that the forum body draws from when Project Ventures awards begin flowing. For AI safety and AI security research groups that are positioning for the next three years of federal funding, the cost of not being in the initial repository is meaningful. The cost of responding is a few weeks of senior-faculty time and a clear-eyed assessment of whether the forum model fits the institution's strategy.

The June 1 announcement is the federal AI safety funding apparatus committing to a specific model of how university-led research will be funded, governed, and transitioned over the next several years. AI Forge is the central instrument. The June 22 RFI is how universities tell the forum body who they are.