IARPA BENGAL — Bias Effects and Notable Generative AI Limitations
Offers super seedling support for large language model threat bias and vulnerability testing toward safer deployment.
IARPA BENGAL, short for Bias Effects and Notable Generative AI Limitations, is a super seedling program under IARPA focused on the threats and vulnerabilities of large language models, including multimodal and text-only systems. It was built for intelligence-community use and centers on tools that can probe, measure, and mitigate model failure modes. The program ran as an 18-month effort beginning in 2026, but the BAA is now closed and no longer accepting new proposals. Six prime performers were selected and began work in early 2026. The available notices do not publish a funding cap, and the earlier competition followed the standard IARPA rules for U.S. organizations, with government agencies, FFRDCs, and UARCs excluded as primes. BENGAL favors teams that can do serious model evaluation rather than general-purpose generative AI demonstrations. The most relevant applicants are those that can build diagnostics, stress tests, or mitigation methods for specific threat domains and then show how those tools can be used in an operational intelligence setting. As a closed competition, it now reads more as a signal of IARPA’s priorities than as an open application route.
No upcoming rounds verified. Cadence: One-off.