Web Application Firewalls (WAFs) are a critical defense against SQL injection (SQLi), yet static rule sets are brittle: they fail to generalize to novel attack variants and degrade under adaptive adversarial pressure. We present a co-evolutionary framework in which Large Language Model (LLM)-powered Attack and Defense agents engage in an iterative adversarial loop. A Hard Payload Feedback mechanism drives adaptation—only payloads that bypass the current defense seed the next generation—forcing the defense to generalize beyond individual attack syntax rather than memorize specific patterns. We compare two rule synthesis strategies: naïve per-payload (one rule per bypass instance) and cluster-based (generalized rules over semantically similar attack families). Evaluated on a custom PHP/MySQL application using GPT-4.1-mini on tautology and conditional SQLi, our results show a consistent advantage for semantic generalization. Under static rules, tautology-based SQLi reaches 100% bypass within three iterations, while conditional SQLi exhibits unstable but high bypass behavior, peaking at 98%. Under co-evolution, naïve per-payload synthesis fails to fully block tautology attacks, accumulating 205 rules over 10 iterations, and achieves full blocking on conditional after growing to 81 rules. In contrast, cluster-based synthesis reaches 0% bypass on both attack families, using only 9 rules for tautology and 3 rules for conditional. These results demonstrate that semantic generalization, rather than reactive patching, is the key to robust WAF hardening.
Data cleaning is a step in a Federated Learning pipeline which attempts to create higher quality data. Papers often evaluate methods in relation to improving utility metrics however they rarely measure the impact on fairness. Our published paper was an experimental report on this issue. This paper presents selected observations from that report which seem most likely to motivate a contribution. We then discuss and evaluate potential contributions.
Electro-mechanical relays are a key component used in Nuclear Power Plant (NPP) safety critical systems. However, Electro-mechanical relays (at the macro-scale) have reliability and maintenance issues, excessive heat dissipation, and large area footprint which tend to limit their use in smaller advanced reactors. In contrast, a configurable architecture built entirely from micro-electromechanical (MEMs) or Nano-electromechanical relays (NEMs) has the potential to effectively eliminate leakage energy losses, operate at high temperatures (175 °C and above) and tolerate radiation doses in excess of 100 Mrad. Motivated by these challenges, this paper presents a novel configurable micro-scale relay architecture that drastically reduces the size of relays [1], enabling thousands of logic functions on a single chip. The proposed architecture provides a foundation for a technology-diverse, low-complexity, radiation-and transienthardened logic architecture suitable for Safety Actuation and Protection Systems in extreme environments [1]–[3].
Modern fingerprint sensors remain critically vulnerable to presentation attacks (PA), as spoofs fabricated from everyday materials can successfully bypass both enrollment and authentication. Our evaluation of seven commercial devices confirms this urgent threat, highlighting the need for robust antispoofing mechanisms beyond software-only countermeasures. To address this challenge, we propose a robust fingerprint recognition system featuring CRISP, a dedicated processor for hardware-anchored liveness detection. By unifying a specialized sensing interface with a near-sensor AI processing engine, CRISP identifies the unique bioelectrical characteristics of live tissue in real time, even in interference-heavy industrial environments, within a single hardware unit. We validate the industrial feasibility of CRISP through non-intrusive parallel integration with a commercial pattern-matching sensor, demonstrating complete end-to-end operation within the Windows Hello application. The integrated system effectively rejects conductive spoofs in 0.82 seconds, providing a secure and reliable anti-spoofing framework for consumer biometric devices.
Malware increasingly conceals its most damaging behaviors, exposing only a controlled surface to sandboxes. To overcome limitations in automated systems, substantial manual effort is required by analysts to reveal hidden behavior in malware. Although Large Language Models (LLMs) offer a new form of assistance, they often misidentify the code that suppresses malicious activity when principled design practices are not applied. Guided by insights distilled from academic research, industry best practices, and feedback from professional malware analysts, we designed LUMEN, an LLM-assisted system that identifies decision points that hide malicious behavior in malware and constructs activation methods to expose them. We showcase LUMEN's performance when analyzing real-world malware and demonstrate its value over traditional sandboxes. To address certain shortcomings of static LLM workflows, we also investigate an agentic architecture that mimics the iterative hypothesistesting behavior of professional reverse engineers.
