The Office of Science and Technology Policy (OSTP) recently invited the public to comment on the 2025 National AI R&D Strategic Plan—not to be confused with the AI Action Plan from earlier this year.

As part of our response to the 2025 National AI Research RFI, we’ve recommended that the U.S. government prioritize investments in quality dataset analysis and adversarial evaluation infrastructure. These capabilities are not only essential for model reliability, but also for discovering emergent behaviors, mapping capability terrain, and stress-testing AI systems in mission-critical or national security contexts. 

Without dedicated funding, testbeds, and scenario libraries for offensive capability exploration, the U.S. risks under-preparing for failure modes that don’t show up in benchmark evals—and ceding ground to adversaries who learn by probing, not by publishing.

View Dreadnode’s recommendations here, or read a summary below.

‍

Increase investments in data science and quality dataset analysis

The foundation of any reliable AI system lies in the quality of its training data. 

Our recommendations emphasize the urgent need for federal R&D investments in dataset contamination detection, bias auditing, and integrity verification tools. We've proposed developing quantifiable benchmarks that can identify hidden biases and contamination issues that directly correlate with AI failure modes like hallucinations and capability degradation. This includes creating evaluation methods that work across both open-source datasets (through direct analysis) and proprietary datasets (through black-box behavioral testing). 

Additionally, we've called for robust post-deployment telemetry standards that can track AI system behavior in real-time, integrating with existing cybersecurity frameworks to enable comprehensive observability. 

A key element of our proposal is establishing a National AI Data Quality Assurance Consortium that would bring together government, industry, and academia to standardize metrics and benchmarks for AI systems deployed in critical infrastructure and government operations.

‍

Promote focused research around AI behavioral anomalies

Understanding how AI systems fail—both through adversarial attacks and natural behavioral drift—is critical for national security applications. 

Our recommendations focus on developing comprehensive red teaming and adversarial testing infrastructure that can detect goal obfuscation, adversarial goal hacking, and agentic manipulation behaviors. This includes testing for non-adversarial anomalies like AI sandbagging (intentional underperformance) and alignment faking (deceptive compliance with human instructions). 

We've proposed that federal funding support software testbeds capable of simulating degraded operational conditions—air-gapped environments, network disruptions, and resource constraints—to evaluate how AI systems behave under stress. Building on the proposed TEST AI Act of 2025, these testbeds would support both model evaluation and the optimization of government AI programs, including systems designed for cybersecurity applications. 

Our vision includes an AI-enabled red teaming consortium that leverages public-private partnerships to systematically stress-test AI capabilities and accelerate vulnerability discovery across the threat landscape.

‍

Gaining confidence in our AI deployments

Maintaining U.S. leadership in AI innovation requires confidence in our AI deployments. This can only be accomplished with additional exploration around dataset integrity, behavioral anomaly detection, and sound testbed infrastructure. 

By strengthening the foundations of AI reliability and enabling secure, scalable deployment across high-stakes domains, we are working toward a future where we can leverage AI's transformative capabilities to the nation's strategic benefit.

Read our complete response for more detailed recommendations. To learn more about Dreadnode’s role in advancing National AI R&D, email daria@dreadnode.io.

‍