Papers

Autonomous AI agents developed using LLMs can reason, invoke tools, browse the web, and execute actions on behalf of users. The very autonomy of such agents creates a direct attack vector linking untrusted external text to side effects in runtime environments. In this paper, we formalize a five-category threat model covering prompt injection, privilege escalation, data exfiltration, destructive execution, and malicious tool invocation. We design a multi-layered software engineering architecture featuring origin tagging, a least-privilege policy engine with typed schemas, human-in-the-loop intervention, and container sandboxing with egress control. Across 1,000 empirical trials, the architecture slashes attack success rates from 48.7% to 4.4% (a 90.9% relative risk reduction) while preserving 88.7% task success.

Investigating non-deterministic race conditions and fault isolation across multi-threaded RTOS kernels (Apache NuttX) and distributed event-driven scalers (KEDA). When high-priority interrupt routines or concurrent state machines race, transient memory corruption leaves zero post-mortem crash telemetry. I am building deterministic execution harnesses that seed pseudo-random scheduler preemption and trace race windows at microsecond precision to guarantee bit-identical fault reproduction: same seed, same crash.