https://gustybear.github.io/author/yingfei-dong/Yingfei DongHugo Blox Builder (https://hugoblox.com)en-usMon, 17 Aug 2026 10:10:00 -1000https://gustybear.github.io/media/logo_hu_d0a0b1783c391ac0.pnghttps://gustybear.github.io/author/yingfei-dong/https://gustybear.github.io/grant/2026_doe_ascr_stratos/Mon, 17 Aug 2026 10:10:00 -1000https://gustybear.github.io/grant/2026_doe_ascr_stratos/<h1 id="executive-summary">Executive Summary</h1> <p>STRATOS develops a cloud-native, model-agnostic security middleware for protecting AI models used in time-critical scientific and energy-system operations. AI is increasingly embedded in workflows such as grid forecasting, contingency analysis, and operational decision making, creating new risks from adversarial perturbations, inference-time evasion, and backdoor attacks that may remain physically plausible while corrupting model outputs. STRATOS addresses this problem through a physics-grounded DevSecOps architecture that continuously evaluates inference streams, computes certified trust scores, and distinguishes malicious manipulation from legitimate operational variability without requiring modification of the protected model.</p> <p>The project combines physics-constrained adversarial emulation, imbalance-resilient certified detection, and a multi-tier runtime mitigation pipeline that includes targeted input purification, Control Barrier Function-based graceful degradation, and quarantine or rollback of compromised data and models. The Phase I system will be evaluated across the University of Hawaiʻi at Mānoa campus microgrid and Argonne National Laboratory&rsquo;s Controller-Hardware-in-the-Loop platform, targeting at least 90% certified detection, no more than 5% false positives, at least 95% interception of synthesized adversarial payloads, and end-to-end latency of 20 ms or less. The longer-term goal is to transition STRATOS toward a federated security architecture for trustworthy AI across DOE scientific computing and critical-infrastructure environments.</p>https://gustybear.github.io/grant/2026_nsf_ge_cyberai/Sat, 15 Aug 2026 15:29:00 -1000https://gustybear.github.io/grant/2026_nsf_ge_cyberai/<h1 id="executive-summary">Executive Summary</h1> <p>AI agents that autonomously browse the web, manage email, execute code, and query databases are being deployed across federal agencies at accelerating pace. These agents create a qualitatively new attack surface: a poisoned email can cause an agent to exfiltrate files; a malicious code comment can redirect an execution pipeline; a hidden webpage instruction can hijack a browsing session. Yet cybersecurity education has not kept pace. Existing AI security tools (Lakera Gandalf, OWASP LLM labs, CTF competitions) teach only chatbot-level prompt injection, a single-turn, text-only model that does not prepare students for agent-mediated threats involving tool access, autonomous action, and multi-step reasoning chains.</p> <p>This project develops <em>AgentSec</em>, a scenario-based educational platform that teaches undergraduate cybersecurity students to identify, exploit, and defend against threats unique to autonomous AI agents, organized into three thrusts. Thrust 1 creates the educational framework, where a trust boundary taxonomy organizes agent vulnerabilities into three progressive categories, i.e., data ingestion, tool-action, and reasoning chain boundaries. This gives students a transferable mental model. Based on the taxonomy, we design a three-module curriculum with six organization-contextualized scenarios, red-team/blue-team experiential learning cycles, and a configurable difficulty framework. Thrust 2 builds the educational technology, where we develop sandboxed environments to enable students&rsquo; interaction with real LLM-powered agents with controlled tool access, a layered assessment system with an AI tutoring agent, an instructor analytics dashboard, and a scenario authoring toolkit for community-contributed content. Thrust 3 evaluates impact through a quasi-experimental study, assessing platform quality, student learning, and workforce placement into government and industry CyberAI roles. All code will be open-source, self-hostable via Docker Compose with open-weight models, and released through CLARK for nationwide adoption.</p>