J. Yeon

Lena Libon, Ben Rank, Jehyeok Yeon et al.

As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor to detect covert sabotage before deployment. We evaluate AI control for automated AI R&D with ResearchArena, a framework spanning four long-horizon tasks: safety post-training, capabilities post-training, CUDA-kernel optimization, and inference-server optimization. Because the deliverable in AI R&D is an artifact that will be deployed and run, we pair each main task with two kinds of hidden side task: an embedded side task that sabotages the submitted model, adapter, kernel, or server itself, and an independent side task that takes a covert out-of-scope action in the sandbox while leaving the artifact honest. We evaluate frontier agents at both sabotage and monitoring, across four monitor types that vary along two axes: whether the monitor may execute and probe the produced artifact, and whether it sees the agent's chain-of-thought. We find that sabotage hidden in the training data is the hardest to catch, flagged fewer than half the time. Letting the monitor run experiments on the artifact, rather than only read the trajectory, helps, but it is not enough: monitors still miss embedded sabotage by inspecting only the surface, by noticing the anomaly but explaining it away, or by probing the artifact with the wrong test. We release ResearchArena as a modular framework for evaluating sabotage and control in automated AI R&D.

Jehyeok Yeon, Hyeonjeong Ha, Qiusi Zhan et al.

Multimodal large language models introduce attack surfaces absent in unimodal systems: adversaries can distribute malicious intent across modalities to evade unimodal safeguards. This motivates using cross-modal consistency as a detection signal rather than inspecting each modality in isolation. Our key observation is that benign inputs induce compatible predictive behavior from text-only and vision-only reasoning that stabilizes when fused, whereas adversarial manipulation disrupts this consistency, causing abnormal multimodal behavior. Existing defenses that examine raw inputs or outputs overlook this internal fusion process, rendering them brittle and computationally expensive. We propose FlowGuard, a lightweight inference-time framework that detects harmful inputs by monitoring internal multimodal consistency. Unlike approaches that rely on scalar confidence metrics, FlowGuard derives FlowVectors inspired by Partial Information Decomposition that quantify cross-modal redundancy, synergy, and modality-specific dominance, capturing whether fused multimodal predictions remain aligned with unimodal semantic evidence. In a one-class classification problem trained solely on benign data, FlowGuard reduces Attack Success Rates from >90% to <15% on unseen attacks, with <3% utility loss and up to a 6 times latency reduction. Our results demonstrate that monitoring cross-modal consistency offers an efficient and effective defense for multimodal reasoning.

Jehyeok Yeon, Ben Rank, Maksym Andriushchenko

AI agents are increasingly used to automate research and development tasks, yet existing benchmarks typically evaluate them on prescribed workflows or narrow action spaces. Even nominally open-ended tasks can often be solved by retrieving a well-known recipe and tuning a few hyperparameters, making it unclear whether strong results reflect genuine optimization or memorized solutions. We introduce InferenceBench, where an agent must deploy an OpenAI-compatible inference server and optimize the speed of LLM inference. Each agent receives a target LLM, one H100 GPU, an optimization scenario, and a wall-clock time budget of two hours. Three optimization scenarios isolate distinct bottlenecks of inference (prefill latency, decode latency, and concurrent request throughput) and a fourth balances all three at the same time. Across 15 frontier agent configurations, agents reliably improve over a naive PyTorch baseline (up to $8.08\times$) and often match or exceed serving engines with default settings ($4.05\times$ for vLLM), but still fall below a simple hyperparameter search under the same time budget (up to $11.53\times$). Qualitative analysis of agent trajectories shows that although agents enumerate many relevant optimization techniques, they overwhelmingly converge on a single inference framework. They test only a few distinct configurations and spend the remaining budget re-measuring, repairing, or optimizing hyperparameters rather than exploring substantially different strategies. This suggests the bottleneck is not domain knowledge, but the ability to propose diverse configurations, evaluate them systematically, and submit the best identified solution. Overall, InferenceBench reflects the ability of agents to operate in an open-ended AI engineering setting, where memorized solutions lead to limited improvements.

Jehyeok Yeon, Isha Chaudhary, Gagandeep Singh

Large language models (LLMs) are increasingly deployed in agentic systems where they map user intents to relevant external tools to fulfill a task. A critical step in this process is tool selection, where a retriever first surfaces candidate tools from a larger pool, after which the LLM selects the most appropriate one. This pipeline presents an underexplored attack surface where errors in selection can lead to severe outcomes like unauthorized data access or denial of service, all without modifying the agent's model or code. While existing evaluations measure task performance in benign settings, they overlook the specific vulnerabilities of the tool selection mechanism under adversarial conditions. To address this gap, we introduce ToolCert, the first statistical framework that formally certifies tool selection robustness. ToolCert models tool selection as a Bernoulli success process and evaluates it against a strong, adaptive attacker who introduces adversarial tools with misleading metadata, and are iteratively refined based on the agent's previous choices. By sampling these adversarial interactions, ToolCert produces a high-confidence lower bound on accuracy, formally quantifying the agent's worst-case performance. Our evaluation with ToolCert uncovers the severe fragility: under attacks injecting deceptive tools or saturating retrieval, the certified accuracy bound drops near zero, an average performance drop of over 60% compared to non-adversarial settings. For attacks targeting the retrieval and selection stages, the certified accuracy bound plummets to less than 20% after just a single round of adversarial adaptation. ToolCert thus reveals previously unexamined security threats inherent to tool selection and provides a principled method to quantify an agent's robustness to such threats, a necessary step for the safe deployment of agentic systems.

Hangoo Kang, Jehyeok Yeon, Gagandeep Singh

Autonomous agentic AI systems powered by vision-language models (VLMs) are rapidly advancing toward real-world deployment, yet their cross-modal reasoning capabilities introduce new attack surfaces for adversarial manipulation that exploit semantic reasoning across modalities. Existing adversarial attacks typically rely on visible pixel perturbations or require privileged model or environment access, making them impractical for stealthy, real-world exploitation. We introduce TRAP, a generative adversarial framework that manipulates the agent's decision-making using diffusion-based semantic injections. Our method combines negative prompt-based degradation with positive semantic optimization, guided by a Siamese semantic network and layout-aware spatial masking. Without requiring access to model internals, TRAP produces visually natural images yet induces consistent selection biases in agentic AI systems. We evaluate TRAP on the Microsoft Common Objects in Context (COCO) dataset, building multi-candidate decision scenarios. Across these scenarios, TRAP achieves a 100% attack success rate on leading models, including LLaVA-34B, Gemma3, and Mistral-3.1, significantly outperforming baselines such as SPSA, Bandit, and standard diffusion approaches. These results expose a critical vulnerability: Autonomous agents can be consistently misled through human-imperceptible cross-modal manipulations. These findings highlight the need for defense strategies beyond pixel-level robustness to address semantic vulnerabilities in cross-modal decision-making.