Strengthening Defense Capabilities Through the Cybersecurity Grant Program

In 2023, OpenAI launched the Cybersecurity Grant Program to support innovative research at the intersection of cybersecurity and AI, and to provide defenders with state-of-the-art AI models. Amid enthusiastic response from the community, the program received over 600 applications and is currently supporting research across a range of fields.

The selected key projects address the following core security challenges:

• LLM Security and Reliability: UC Berkeley's Wagner Lab is researching defense techniques against prompt-injection attacks, while Mithril Security is advancing inference infrastructure hardening using TPM-based secure enclaves to deploy AI models without exposing data.
• Code and Configuration Security: Boston University's SeclaBU is researching LLM-based code vulnerability detection and remediation, while Coguard is developing technology that uses AI to automatically detect software misconfigurations and prevent security incidents.
• Autonomous Defense and Red Teaming: UCSC's CY-PHY Security Lab is exploring the design of autonomous cyber defense agents using foundation models, while MIT CSAIL is automating red-teaming activities through prompt engineering and testing the capabilities of LLM agents in CTF environments.
• Data Privacy and Education: Dartmouth's Breuer Lab is developing techniques to defend against model training data reconstruction attacks, while Gabriel Bernadett-Shapiro provides technical education to experts and students through AI OSINT workshops and security starter kits.