AI Forge Cycle 2: Up to 15 Universities, $1.5M Each, and a Two-Page Abstract Due November 6
November 3, 2026 · 7 min read
Granted Research Team · Editorial policy
The second round of AI Forge, the university research forum that DARPA and the National Science Foundation set up in June, is now open, and its first deadline is close. Riverside Research, the nonprofit that administers the forum, released the Cycle 2 Request for Abstracts through the GovSignals portal on October 23, 2026. Abstracts are due at 11:59 PM ET on Thursday, November 6. Invited teams then submit a pitch and a full proposal by December 8 and present in person at a Live Pitch Day in Washington, D.C., on December 11.
The award terms are specific. Per the RFA terms relayed in university limited-submission notices, DARPA will select up to 15 universities. Each Project Venture is capped at $1.5 million, covering labor, travel, subcontracts and overhead, plus a separate compute and model-access budget of up to $1 million. The period of performance is 12 months, split into two six-month phases, and no cost share is required. For a forum that earlier described its ventures as "$750K to $3M or higher," Cycle 2 sets a firmer ceiling.
This piece covers what changed between Cycle 1 and Cycle 2, the eligibility rules that will knock out applicants this week, and how to write a two-page abstract that gets past a down-select designed to be fast.
What AI Forge Is, and What It Is Not
AI Forge is not a standard NSF solicitation or a DARPA Broad Agency Announcement. DARPA announced it on June 1, 2026 with NSF and NIST's standards center, now called the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI). It runs as a forum: government mission stakeholders and frontier AI companies set the research agenda, a nonprofit administers it, and universities compete for short, milestone-driven Project Ventures. Program manager Matthew Marge sits in DARPA's Information Processing Techniques Office. The June request for information (DARPA-SN-26-80) drew nearly 130 responses, according to DARPA.
If you followed the June launch through our earlier coverage of the AI Forge RFI and Project Ventures, one change is visible right away: DARPA's AI Forge program page now redirects to "SI Forge." A footnote in the Cycle 2 challenge report explains that the report was written before Executive Order 14434, which directs the government's use of the term "Super Intelligence," and that DARPA is updating its materials. The same rename has run through federal science programs this fall, including NSF's SINI instrument program and NOAA's climate machine-learning call. Your abstract can use either term, but expect reviewers and any future notices to say "SI."
The Cycle 2 Agenda: Four Thrusts, Nine Challenges
The 21-page Critical AI Challenges for National Security, Cycle 2 report, released October 21, is the technical basis for the RFA. It was built with input from government mission stakeholders and frontier AI companies, and it cites two policy anchors: America's AI Action Plan (July 2025), which tells DARPA, CAISSI and NSF to advance interpretability, control and robustness, and National Security Presidential Memorandum 11 (June 5, 2026), which calls for "reliable, robust, steerable, and controllable" mission AI under an accountable chain of command.
| Thrust | Challenges | Capability question |
|---|---|---|
| 1. AI Interpretability | 1.1 Predict Emerging Failure; 1.2 Establish Causal Accountability | Can we detect emerging failures early enough to act, and establish their cause and correction? |
| 2. AI Control | 2.1 Enforce Authorized Bounds; 2.2 Verify Steerability as Objectives and Conditions Change; 2.3 Control Autonomous R&D and Experimentation | Can we keep effective authority over agents as objectives change and capabilities grow? |
| 3. Adversarial Robustness | 3.1 Preserve Mission Integrity Against Adaptive Adversaries; 3.2 Sustain Secure and Available AI Through Compromise and Disruption | Can we preserve mission capability as adversaries adapt, and restore secure operations afterward? |
| 4. AI Assurance (new) | 4.1 Verify Scientific Outputs of AI Systems that Exceed Human Expertise; 4.2 Maintain Defensible Mission Assurance Over Time | Can we verify difficult AI results and justify continued trust as conditions change? |
Two changes from Cycle 1 matter most.
A fourth thrust. Cycle 1 had three thrusts: interpretability, control and adversarial robustness. Cycle 2 adds AI Assurance, which covers evaluation, verification and the evidence needed to decide whether a deployed system should keep running. Challenge 4.1 is the most unusual item in the report. It asks how to verify scientific outputs from AI systems that exceed human expertise, where no ground truth exists to check against. Research groups in formal methods, statistics and metascience that would not have seen themselves in Cycle 1 now have a direct entry point.
Agents, not models. The report says Cycle 2 "emphasizes increasingly autonomous agents that operate over time with tools, persistent memory, delegated tasks, and changing conditions." It also argues that "a model evaluation cannot establish such properties for a complete mission system." Errors build up through plans, memory, tool calls and delegated work, even when each individual action looks fine. Teams must specify the complete agent configuration: models, persistent context, memory, tools, permissions, subagents and external services. Multi-agent demonstrations should vary group size, membership, communication delays and task allocation. Several challenges ask for methods that work across systems "assembled from different developers' models, tools, and services, including when access to model internals is limited." An abstract built around a single open-weights model with full white-box access will read as narrower than what DARPA wants.
The Evidence Bar
The report's closing section, "Research Execution and Evidence of Breakthrough," tells reviewers what to score. Each venture should establish "a reproducible capability, principle, or limit." Teams must compare against the strongest published, reproducible baselines, with model capability, compute and operator involvement matched where feasible. They must also show what their work adds "beyond assembling existing monitors, access controls, defenses, or benchmarks."
Three details in that section are easy to miss:
- Negative results count. "Results that reveal a reproducible limitation are valuable when they identify conditions under which a widely used approach fails." A credible plan to break a popular monitoring or red-teaming method is a legitimate venture.
- Fixed benchmarks are not enough. Wherever an adversary or strategic agent is involved, it must be allowed to adapt to the actual defense under stated limits. Benchmarks "cannot alone establish effectiveness under adaptation."
- Go/no-go evidence. Demonstrations should support go/no-go decisions against project-specific criteria tied to mission consequences. With a 6-Month Review Day built into the 12-month award, Phase 1 needs a result DARPA can judge at month six.
DARPA also says the agenda will be reviewed every six months. Cycle 3 will likely retire resolved questions and add new ones, so a near-miss this round is not wasted.
Eligibility Rules That Will Stop Applications This Week
University limited-submission notices from Stony Brook and the University of Colorado Boulder spell out several constraints:
- One abstract per institution in practice. Both schools are treating the call as a limited submission, and internal deadlines have already passed (CU's on October 29, Stony Brook's on October 26). If your research office has not chosen its nominee, it is very late to start.
- One abstract per PI, and each abstract may address only one challenge.
- PI effort of at least 20 percent.
- Key personnel citizenship. Stony Brook's notice says PIs and co-PIs must be U.S. citizens or permanent residents, and that foreign nationals need a Foreign National Justification form if the abstract advances. Confirm the exact rule in the RFA before you list your team.
- In-person Pitch Day. A PI who cannot be in Washington on December 11 is excluded.
- No pre-award costs. Spending before award cannot be recovered.
The RFA also lists what DARPA will not fund: work that duplicates frontier-lab research, topics industry is likely to solve within 6 to 12 months, broad claims without technical detail, and AI-generated abstracts. DARPA called the AI-generated abstracts it received in Cycle 1 "weak."
Finally, the intellectual property terms. Venture IP is meant to be shared among forum participants, preferably under open-source licenses such as MIT. Applicants must also distinguish their project from active industry-partner projects. Tech transfer offices that count on licensing revenue should read these terms before an abstract goes out, not after an invitation to pitch.
How to Use the Three Days Before November 6
Lead with the required sentence. The abstract opens with a one-sentence objective in a fixed format: "[Project Title] will develop [technology] to enable [capability]." Write it first. If you cannot fill in "capability" with something a mission operator would recognize, the challenge fit is not there yet.
Choose the challenge where your evidence is strongest, not the one that sounds most impressive. With nine challenges and one allowed per abstract, the temptation is to stretch across thrusts. Pick one and cite its driving question word for word.
Make the six-month result concrete. Name the Phase 1 deliverable DARPA will see at the review: a measured warning lead time on held-out failure types for Challenge 1.1, or a demonstrated bound violation rate under an adaptive attacker for Challenge 3.1. Vague phasing reads as the "broad claims" DARPA screens out.
Budget compute separately, with contingencies. The compute estimate is a separate one-page document, broken out by phase, with resource types, quantities, unit prices, preferred frontier models and acceptable alternatives. List any free or discounted university resources. Plan Phase 2 for both low and high compute availability. Exceeding the $1 million compute ceiling requires technical justification and depends on progress.
Write it yourselves. The RFA says DARPA noticed machine-written abstracts in Cycle 1 and rated them poorly. Specific claims, real baselines and your own prior results are the best way to avoid sounding generic.
What It Means
AI Forge is DARPA's attempt to aim academic AI safety research at deployed, multi-vendor agent systems, faster than the usual grant cycle allows. With one abstract per university, up to 15 awards, and about five weeks from abstract to Pitch Day, the bar is high and the schedule is short. Universities that miss Cycle 2 should still treat the challenge report as a guide to what DARPA and NSF will fund in this area for the next year, and build preliminary evidence against it now.
If your team is advancing to the December 8 pitch and full proposal, Granted can help you turn the abstract into a full proposal that matches the challenge report's evidence requirements.