The paper introduces PoisonForge, a comprehensive benchmark demonstrating that even a small number of targeted poisoned examples can significantly compromise the safety and reliability of instruction-…

The paper introduces a kill-chain canary methodology to diagnose prompt injection vulnerabilities across multi-stage LLM pipelines, revealing that write-node placement and document format are critical…

The paper introduces the Task Alignment Benchmark (TAB) to evaluate terminal agents' ability to selectively follow relevant environmental instructions while ignoring misleading distractors, revealing…

This paper investigates indirect prompt injection vulnerabilities in ReAct agents by systematically analyzing how the injection depth and payload framing affect attack success rates, finding that inje…

The paper investigates indirect prompt injection vulnerabilities in ReAct agents by systematically varying the injection depth, payload framing, and turn budget, finding that injection depth is the do…

The paper introduces Symbolicate-Enrich-Sample, a pipeline that efficiently filters millions of functions in a Windows OS to create a highly prioritized, manageable shortlist of potential vulnerabilit…

The paper introduces Symbolicate-Enrich-Sample, a low-cost pipeline that drastically reduces the search space of a whole operating system by prioritizing vulnerable functions, turning millions of pote…

The paper introduces TRAILS~, a novel method that improves code correctness validation by grounding LLM reasoning in concrete (input, output) pairs derived from specifications, achieving state-of-the-…

The paper presents Symbolon, a framework that learns and applies context-sensitively diverse code transformations to improve symbolic execution, increasing coverage and reducing costs.

The paper proposes a comprehensive application-layer reference monitor to detect and mitigate data exfiltration via covert channels embedded in LLM agent egress payloads across text, image, and audio…

This paper proposes a portable inlining-prediction framework for lightweight systems, using production compiler diagnostics, an extractor, and a trained predictor.

FuzzPilot is a controller for AFL++ that validates candidate mutation recipes by running short micro-campaigns, demonstrating a mechanism to manage fuzzing plateaus, though initial results on a satura…

The vulnerability of LLM agents to prompt injection depends not on the specific channel (tool output vs. tool description) but on the interaction between the model and the surface.

The vulnerability of LLM agents to prompt injection depends not on the specific channel (tool output vs. tool description) but on the interaction between the model and the surface itself.

The paper proposes Fuzzy-Function Programming and introduces Program-as-Weights (PAW), a compact, locally-executable neural artifact for everyday programming tasks.

The paper presents an initial implementation of a quantized, integer-only transformer for jet tagging on the AMD Versal AI Engine using a reusable software framework.

This paper proposes a new training criterion to reduce a classifier's reliance on shortcuts in language proficiency assessment systems, improving their correlation with human references.