Agents in Research
« Agents in Research

#21 — Volumetric imaging adopts AI agents, MetaboT framework

August 12, 2026
A look at the shift to multimodal, agent-driven AI systems in volumetric medical imaging for biomarker discovery and adaptive clinical workflows. Further applications of agentic AI include autonomous quantum sensing experiments and natural language querying of metabolomics knowledge graphs. Also covered are security implications, such as intellectual property leakage in autonomous multi-agent systems and the changing landscape of offensive security.

Volumetric medical imaging transitioned from classical deep learning to multimodal, agent-driven AI systems. These new systems identify volumetric biomarkers and use diverse data for predictive and generative modeling. AI agents and foundation models now provide advanced reasoning and adaptive clinical workflows in volumetric medical imaging.

MetaboT, an open-source multi-agent Large Language Model (LLM) framework, translates natural language questions into executable SPARQL queries over metabolomics knowledge graphs. Its modular architecture addresses issues like hallucination and schema compliance from single-model approaches. Validated on the Experimental Natural Products Knowledge Graph (ENPKG), MetaboT achieved 83.67% accuracy on a metabolomics benchmark, compared to 8.16% for a single-shot baseline. The framework reduces technical barriers for metabolomics researchers, enabling semantic data mining without specialized programming.

Researchers introduced Agent Harness Distillation (AHD) to analyze security risks, specifically inference-time harness extraction from Autonomous Multi-Agent Systems (AMAS). Experiments confirmed AHD's effectiveness and demonstrated intellectual property leakage risks in AMAS. This research formalized harness extraction as a new security problem. A deception-based defense was proposed to mitigate these risks.

An agentic AI workflow was implemented for autonomous experiments on nitrogen-vacancy centers in diamond. This workflow combines project records, quantitative calculation, data analysis tools, and experiment control. It performed tasks such as selecting an NV center, calibrating its resonant frequency, and conducting specific measurements. Two offline benchmarks were introduced to evaluate AI agent reasoning independently of laboratory execution: a Ramsey checkpoint benchmark and a pulsed optically detected magnetic resonance (pODMR) data evaluation benchmark. These benchmarks were evaluated using GPT-5.4, GPT-5.5, and GPT-5.6 Sol models. Greater reasoning effort improved recognition of residual resonance calibration offset in the Ramsey checkpoint benchmark. In the pODMR benchmark, pulse sequence information alone increased false positive resonance judgments at higher reasoning effort; however, incorporating an expected signal calculation maintained low false positive rates across all models and reasoning settings. Research results suggested a division of labor for autonomous experiments: AI agents generate scientific hypotheses and analyze data using quantitative tools, while deterministic code controls hardware and enforces safety protocols.

LLM-driven autonomous agents are changing offensive security. They exhibit non-deterministic actions, open-ended impacts, and a reduced skill floor for users, differing from traditional penetration testing methods. This combination facilitates the industrialization of offensive capabilities, initially favoring attackers. These characteristics also present challenges for existing cybersecurity and AI-ethics frameworks, complicating moral attribution.

Sources

  1. A Comprehensive Review Tracing the Evolution of Volumetric Medical Imaging Analysis from Classic CNNs to Emerging AI-Agents
    Abstract Volumetric medical imaging has redefined modern healthcare, enabling precise diagnosis, prognosis, and treatment planning. During the past decade, the field has undergone a paradigm shift from classical deep learning architectures to multimodal, agent-driven AI systems capable of uncovering rich volumetric biomarkers and utilizing heterogeneous data for predictive and generative modeling. Existing surveys are fragmented, focusing on specific models or tasks instead of offering a unified view of volumetric learning evolution. This study traces the evolution from classical models…
  2. MetaboT: an LLM-based multi-agent framework for interactive analysis of mass spectrometry metabolomics knowledge graphs
    Mass spectrometry-based metabolomics generates complex, high-dimensional data that holds vast potential for biological discovery but remains difficult to integrate and interpret. Knowledge graphs (KGs) unify this heterogeneous information by representing spectra, annotations, taxa, chemical classes, and biological activities as a single interoperable network; however, their practical use is limited by the steep learning curve of corresponding specialized representation and query languages. Here we introduce MetaboT, an open-source multi-agent Large Language Model (LLM) framework that…
  3. Agent Harness Distillation: Inference-Time Harness Extraction and Exploitation in Autonomous Multi-Agent Systems
    Autonomous multi-agent systems (AMAS) built on large language models (LLMs), such as Hermes, increasingly rely on inference-time harnesses to coordinate reasoning and action. Constructing these harnesses requires substantial engineering effort and computational resources, as they are iteratively optimized over a combinatorial search space while co-evolving with the underlying LLM. Inference-time harnesses therefore constitute valuable intellectual property (IP). Although prior work has investigated IP leakage in static multi-agent systems with pre-configured architectures, it remains unclear…
  4. Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
    We implement an agentic AI workflow built around a large language model (LLM) agent for autonomous experiments with nitrogen-vacancy (NV) centers in diamond. NV centers are a widely used platform for quantum sensing, and the ability to control many measurements from a computer makes NV experiments a natural setting for autonomous workflows. We make two main contributions. First, we demonstrate an autonomous NV experiment workflow that combines persistent project records, quantitative calculation and data analysis tools, and deterministic experiment control. In one autonomous experiment, the…
  5. The Ethics of Autonomous AI Agents for Offensive Security
    LLM-driven autonomous agents are reshaping offensive security. Unlike traditional penetration-testing tooling - deterministic, narrowly scoped, and operated by trained practitioners - agentic security tools exhibit indeterminacy along three independent dimensions. First, their actions are drawn from a non-deterministic policy whose outputs resist both ex-ante and ex-post explanation. This complicates incident attribution and pre-deployment safety reviews. Second, their impact is open-ended due to their non-deterministic actions, agency of utilized models, and opaque LLM supply-chains. Third,…

Also this week

Full transcript
AI agents are changing volumetric medical imaging from a diagnostic tool into an adaptive partner in clinical workflows. This is the kind of developing story we track on Agents in Research. We start there. A review of volumetric medical imaging analysis shows a clear trend over the last decade. We're seeing a move away from classical deep learning models. And a move toward what? Toward multimodal, agent-driven AI systems. These are systems capable of discovering volumetric biomarkers on their own and using very different types of data for predictive modeling. So they have more advanced reasoning capabilities. It sounds like they're less of a static tool and more of an adaptive partner in a clinical workflow. Exactly. There's a specific example of this in biological data with a new open-source framework called MetaboT. It's designed for metabolomics research. What does it do? It translates questions, asked in natural language, into executable SPARQL queries that can run on metabolomics knowledge graphs. So instead of a researcher needing to know how to write a complex query, they can just ask a question. Why is this a multi-agent system, though? Why not just use one large model? Because single-model approaches have limitations, like a tendency to hallucinate answers. MetaboT uses a modular architecture with specialized agents. One validates the scope of the question, another generates the query, a third interprets the results. And does that specialization actually work better? How does it perform? Researchers tested it on an expert-authored benchmark. The MetaboT framework achieved an accuracy of over 83 percent. Compared to what for the single-model approach? Just over 8 percent. The multi-agent framework also reduced schema-mismatch errors. The whole point is to lower the technical barrier for researchers. So that's biological research. You mentioned this is happening elsewhere. What about in the physical sciences? There, a research team has put an agentic AI workflow into practice for conducting autonomous experiments on nitrogen-vacancy, or NV, centers in diamond. Autonomous experiments? So the AI is running the lab? To an extent. The system is built around a large language model agent that integrates project records, calculation tools, and experiment control. It demonstrated it could autonomously complete tasks like selecting an NV center and performing standard measurements on it. How do you even evaluate that? It seems hard to separate the agent's reasoning ability from its ability to just execute a lab procedure. That's a key question. The team also introduced two offline benchmarks to do just that. One is a Ramsey checkpoint benchmark, and the other evaluates pulsed optically detected magnetic resonance data. They tested these using several GPT models. And what did they find? They found that giving the agent more reasoning effort improved its ability to recognize certain calibration offsets. They also saw that just giving it raw pulse sequence information led to more false positive judgments. But if they gave it more context? Correct. When they incorporated an expected signal calculation, the false positive rates stayed low. The results suggest a specific division of labor for these autonomous experiments. Where the AI isn't running everything, but handling specific tasks. What's the breakdown? The AI agent handles hypothesis generation and data analysis, while deterministic code is used for the direct hardware control and the safety protocols. Which makes sense. But as these systems get more capable, the security questions get louder. What about the risk of the technology itself being stolen? That's being addressed. A new framework called Agent Harness Distillation, or AHD, was developed to analyze the security risk of extracting these inference-time harnesses from multi-agent systems. It formalizes this kind of intellectual property leakage as a new security problem. So has this been shown to be a real threat? Yes, experiments on real-world systems confirmed the AHD framework is effective and demonstrated that the risk of IP leakage is substantial. To counter it, researchers have proposed a deception-based defense to make extraction harder without hurting the agent's performance. This connects to another area of security research—using these LLM-driven agents for offensive security, like penetration testing. Right, and that introduces what researchers call three dimensions of indeterminacy. The agent's actions aren't always deterministic, their impacts are open-ended, and they lower the skill floor for users to conduct these operations. Which complicates everything. It makes it much harder to attribute moral responsibility and doesn't fit well with existing frameworks for dual-use tech or AI ethics. And this is all part of a much broader series of developments. A new book was just published on ethical considerations for AI developers, and a separate study identified themes for AI adoption in Nigerian commercial banks. We're also seeing new conceptual tools, like a framework for human-centric AI integration, and papers arguing opposite sides—one saying AI strengthens functions in global problem-solving, and another identifying the risks it creates in those same environments. It’s a field trying to figure itself out in real time. Another paper even proposed a model called 'augmented diplomacy' to integrate AI with human judgment. It seems that across science, security, and policy, the capabilities are moving quickly, while the frameworks for managing them are all racing to catch up. We'll have more on the application of AI agents in science next time. From Agents in Research, thanks for listening.