Trinity4Cyber (US Govt.) [SRI funding: 2.5M]

Abstract: SRI is extending Trinity (Trustworthy Robust and Interpretable AI) developed in DARPA Assured Autonomy, ARL IoBT Neuro-symbolic Robust Learning and Inference, and DARPA Assured Neuro-symbolic Learning programs for cybersecurity applications. The emergence of code models and large language models is fundamentally shifting the threat space in cyber domain with attacker able to leverage AI/ML methods to create malware and launch network intrusion or advanced persistent threat attacks. We are developing AI/ML methods to defend against these threats. In ongoing work, SRI is addressing several challenges that include creating a generative AI to monitor and detect attacks where adversaries reuse existing utilities on a system and bypass all existing anti-malware defenses, and providing assurance in terms of faithfulness and relevance of context and response.