Synthesizing Multi-Agent Harnesses for Vulnerability Discovery

GPUBreach demonstrates a fundamentally new class of security vulnerability—GPU-side privilege escalation via Rowhammer that crosses the GPU-CPU boundary, defeating IOMMU protections and achieving root shell access. This opens an entirely new attack surface with broad implications for GPU security architecture, cloud computing, and hardware design. While Paper 2's Chrome zero-days are impressive practical results, its contribution is primarily engineering (harness optimization for LLM agents), building on existing agent frameworks. GPUBreach's novelty in bridging GPU Rowhammer to full system compromise represents a deeper, more foundational security insight with lasting architectural implications.

While Paper 1 presents a robust framework for AI agent safety, Paper 2 demonstrates exceptional real-world impact by discovering ten zero-day vulnerabilities in Google Chrome, including two critical sandbox escapes. Paper 2's introduction of a typed graph DSL and feedback-driven optimization loop represents a major leap in automated vulnerability discovery. The tangible results (CVEs in a major browser) and state-of-the-art benchmark performance will drive massive immediate attention, industry adoption, and citations across cybersecurity and artificial intelligence communities, giving it significantly higher scientific and practical impact.

Paper 2 demonstrates profound real-world impact by discovering ten previously unknown zero-day vulnerabilities in Google Chrome, including two critical CVEs. While Paper 1 offers strong theoretical and empirical advancements in blockchain scalability, Paper 2 tackles a highly timely problem (LLM agents for cybersecurity) with an innovative feedback-driven synthesis approach. Automatically discovering critical vulnerabilities in heavily audited, widely-used software represents a massive breakthrough with broad implications for both software security and AI-driven automated reasoning. This exceptional concrete evidence of real-world applicability gives Paper 2 a substantially higher potential scientific impact.

Paper 2 introduces a fundamentally novel attack vector by exploiting hardware-specific floating-point divergence to trigger LLM backdoors. While Paper 1 shows exceptional practical results (Chrome zero-days) via agent optimization, Paper 2 exposes a foundational blind spot in AI auditing and model supply chains. This conceptual leap—using hardware math variations as an input-independent trigger—creates an entirely new subfield spanning AI security, robustness, and hardware-software co-design, indicating a broader and more enduring scientific impact.

Paper 2 demonstrates higher potential scientific impact due to its concrete, verified real-world results: discovering 10 zero-day vulnerabilities in Google Chrome, including two critical sandbox escapes with assigned CVEs. This provides undeniable evidence of practical impact in cybersecurity. The AgentFlow DSL and feedback-driven harness optimization framework also introduces a novel, generalizable methodology for multi-agent system design. While Paper 1 (Bifrost) presents a solid engineering contribution combining TEE and FHE for private LLM inference with impressive speedups, it remains largely an architectural optimization with estimator-based comparisons rather than demonstrated real-world deployment. Paper 2's breadth of impact across security, AI agents, and software engineering, combined with its immediate practical significance, gives it the edge.

Paper 2 has higher impact potential: it introduces a broadly applicable, typed DSL + feedback-driven search framework for synthesizing multi-agent harnesses, and demonstrates unusually strong real-world validation by discovering multiple previously unknown Chrome zero-days (including critical sandbox escapes). This indicates direct practical utility and high timeliness for AI-assisted security auditing. While Paper 1 is novel and efficient (inference-time activation/logit biasing) and likely influential for secure code generation, its contributions are more incremental and narrower in application scope than an end-to-end system that measurably improves vulnerability discovery in flagship software.

Paper 2 demonstrates exceptional real-world impact by discovering previously unknown critical zero-day vulnerabilities in a major software project (Google Chrome). Its approach to automating multi-agent LLM harnesses addresses a highly timely and critical problem in cybersecurity. While Paper 1 offers strong architectural improvements for FHE, Paper 2's immediate applicability, state-of-the-art benchmark performance, and novel synthesis of LLM agents with runtime feedback represent a broader and more urgently impactful advancement across both AI and security domains.

While Paper 1 provides a valuable benchmark for AI exploitation capabilities, Paper 2 demonstrates profound real-world impact by discovering 10 previously unknown zero-day vulnerabilities in Google Chrome, including two critical sandbox escapes. The introduction of a novel, feedback-driven DSL for synthesizing agent harnesses represents a significant methodological advancement with immediate, high-stakes applications in proactive vulnerability discovery.

Paper 1 audits a privacy system deployed on billions of Apple devices, revealing that differential privacy guarantees are violated for the majority of data collection. This has immediate, massive real-world impact on user privacy and will likely force Apple to fix critical bugs. It also advances the methodology for auditing proprietary privacy systems. Paper 2 is impressive with its Chrome zero-days, but its impact is more incremental within the AI-for-security space. Paper 1's findings affect a far larger population and address fundamental trust in deployed privacy claims, with broader implications for regulatory accountability and corporate transparency.

Paper 1 provides a comprehensive systematization of knowledge (SoK) on reconstruction attacks against synthetic tabular data, contributing a taxonomy, 14 attacks across 9 methods, novel methodology for interpreting attack success, and externally validated results (winning NIST CRC). Its breadth of impact spans privacy, data governance, and policy, influencing how synthetic data is deployed across many domains. Paper 2 is impressive with real zero-day Chrome vulnerabilities, but is narrower in scope (LLM agent harness optimization for vulnerability discovery) and its methodological contributions (a DSL and feedback loop) are more incremental relative to Paper 1's foundational systematization of an increasingly critical privacy problem.