An agentic AI framework developed at Oak Ridge National Laboratory is accelerating plant phenotyping by converting a manual, weeks-long analysis process into an interactive loop measured in seconds. This development is part of a broader look at AI in research, including a new chronological system for tracking multi-agent provenance and ensuring accountability. The summary also covers identified security vulnerabilities in enterprise GenAI tools and the corresponding need for comprehensive governance strategies. Finally, recent studies evaluate how current AI frameworks perform on real-world scientific tasks and propose new methods for addressing cognitive security risks from LLM outputs.
Oak Ridge National Laboratory developed an agentic AI framework for high-throughput plant phenotyping data analysis. This system converts the facility into an interactive, autonomous discovery platform. It features a conversational Co-Scientist Agent for natural language interaction and a Compute Agent on the Frontier exascale supercomputer, communicating securely. This framework reduced analysis time from days/weeks to seconds, enabling agents to suggest analyses and respond to inquiries rapidly.
A study proposed a chronological system for tracking content origin in multi-agent generative chains. This system uses signed, time-stamped symbolic chronicles, updated with each generative step, to attribute contributions.
Research evaluated five AI frameworks (including Kosmos, K-Dense) on scientific tasks such as molecular property prediction. These frameworks showed capabilities for generating hypotheses and testing studies. Human domain expertise remained essential for verifying AI outputs. The study concluded AI frameworks are useful for research prototyping and stress-testing existing studies, noting many limitations are addressable.
Research into GenAI tools like Microsoft 365 Copilot identified security vulnerabilities, including AI-based prompt injection, unintended data synthesis, and cross-tenant data exposure. A comprehensive governance strategy, encompassing technical, organizational, and regulatory controls, is necessary to manage these risks while maintaining GenAI productivity. Existing security measures were deemed insufficient, with security risks common across AI-augmented SaaS platforms. Additionally, a framework called Cognitive Security for LLMs (CogSLLaM) was introduced to address risks from erroneous or misleading LLM outputs that can affect user judgment.
Automated workflows have demonstrated large-scale mobilization of biodiversity data. These systems converted content from 198 journals, releasing structured data from over 95,000 publications, including taxonomic treatments, figures, and material citations.
Sources
- An Agentic AI Framework to Accelerate Scientific Discovery in Plant Phenotyping
High-throughput plant phenotyping now generates image derived datasets far faster than scientists can analyze them. At Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory (APPL), automated stations image hundreds of plants daily across multiple remote sensing modalities; yet, trait extraction and interpretation remain manual, expert-bound, and strictly post-hoc, making analysis, not acquisition, the binding constraint on discovery. We present an end-to-end agentic AI framework that turns the facility from a data factory into an interactive autonomous, discovery platform,…
- Chronology of multi-agent interactions for provenance of evolving information
Abstract Provenance is the chronological history of things, resonating with the fundamental pursuit to uncover origins, trace connections and situate entities within the flow of space and time. As artificial intelligence (AI) advances towards autonomous agents capable of interactive collaboration on complex tasks, the provenance of generated content becomes entangled in the interplay of collective creation, where contributions are continuously revised, extended or overwritten. In a multi-agent generative chain, content undergoes successive transformations, often leaving little, if any, trace…
- Security and Privacy Implications of Microsoft 365 Copilot and GenAI Integration in Enterprise Environments
The introduction of Generative Artificial Intelligence (GenAI) into the world of enterprise Software-as-a-Service (SaaS) is a crucial development in digital transformation. As one of the highest-performing applications of large language models (LLMs) to productivity software, Microsoft 365 Copilot is a product that promises productivity-level efficiency to increase but at the same time presents a difficult compound pose on the security and privacy issue. The paper has used a secondary qualitative study to review the threat climate of Copilot, privacy-leakage threats, and governance. The study…
- CogSLLaM: Cognitive Security for Large Language Models
Large Language Models (LLMs) are increasingly employed across general purpose, expert, and domain-intensive user contexts; in all instances, users rely on these systems for information summarization, task planning, and problem-solving, and even for advice and recommendations related to decision-making. LLMs are becoming embedded in highly technical do-mains, supporting cyber security, code development, health care, organizational policy analysis, and the interpretation, specification, and implementation of complex technical requirements. This expanding range of applications underscores the…
- The Disentis Roadmap: a decadal roadmap for liberating global biodiversity knowledge from scientific literature
A vast share of biodiversity knowledge remains inaccessible, embedded in traditional literature and fragmented electronic resources. This limits scientific progress, constrains evidence-based policy, and hinders full use of rapidly expanding digitised data from natural history collections, citizen science, monitoring programmes including eDNA, and environmental impact assessments. Although biodiversity information is inherently well-structured and suitable for training emerging AI systems, current bottlenecks hinder its integration into Large Language Models and other automated tools.…
- The Disentis Roadmap: a decadal roadmap for liberating global biodiversity knowledge from scientific literature
A vast share of biodiversity knowledge remains inaccessible, embedded in traditional literature and fragmented electronic resources. This limits scientific progress, constrains evidence-based policy, and hinders full use of rapidly expanding digitised data from natural history collections, citizen science, monitoring programmes including eDNA, and environmental impact assessments. Although biodiversity information is inherently well-structured and suitable for training emerging AI systems, current bottlenecks hinder its integration into Large Language Models and other automated tools.…
- Real Science Is Harder Than Benchmarks: Evaluating Advanced AI Frameworks on Published Studies. I. Uncertainty Quantification, ML on Therapeutic Data Commons, and Agent-Based Modeling
Abstract Artificial Intelligence (AI) frameworks for automating scientific research have shown strong performance on benchmarks, but their capacity to routinely reproduce results from multiple real-life published studies remains largely untested. We evaluated five advanced AI research frameworks (Kosmos, K-Dense, ToolUniverse, BioAgents from bio.xyz, and the AI Scientist-v2 from Sakana AI) on three real-life tasks (including two recently published papers) spanning uncertainty quantification for molecular property predictions, machine learning on Therapeutic Data Commons benchmarks, and…
Also this week
Full transcript
A manual analysis of plant data that took weeks is now an interactive process measured in seconds. That change is driven by agentic AI, a topic we cover on Agents in Research. Today, we begin with a framework from Oak Ridge National Laboratory.
New work is showing how AI agents are being applied to speed up data analysis, particularly in scientific fields.
Which has been a huge bottleneck. Scientists can generate these massive datasets with high-throughput imaging, but then analyzing that data can take longer than collecting it.
Exactly. And researchers at Oak Ridge National Laboratory have developed a framework to tackle that for plant phenotyping. They’ve essentially turned their lab into an autonomous discovery platform.
How does that work in practice? Is a scientist just typing questions into a prompt?
Pretty much. There's a conversational 'Co-Scientist Agent' that takes natural-language questions and turns them into analysis plans. Then a separate 'Compute Agent' executes those tasks—like image segmentation—on the Frontier exascale supercomputer.
And it's the separation between those two agents that's notable. They operate in different security domains and communicate through a secure channel. That seems to address two issues at once: security and data provenance, knowing where the data came from and how it was handled.
It does. And the result is that a process that used to take days or even weeks now happens in seconds, in an interactive loop where the agents can reason over the results and respond to the scientist.
We're seeing a similar application for mobilizing biodiversity information from scientific literature. Automated workflows are pulling structured data from over 95,000 publications.
And the scale is massive—760,000 taxonomic treatments and 1.7 million material citations released as FAIR data, meaning it's findable, accessible, and reusable.
So as these frameworks get more common, the next logical question is: how well do they actually perform? They're fast, but are they right?
Right. A recent study evaluated five of these advanced AI frameworks—names like Kosmos, BioAgents, and AI Scientist-v2—on real-world scientific tasks, like molecular property prediction.
The conclusion seems to be that they're useful for initial prototyping or for stress-testing existing studies, but they can't operate alone. The study found that human domain expertise was essential for verifying the AI's output.
Which creates its own challenge. If you need a human to verify the work, you also need to be able to track where every piece of generated content came from, especially in a system where multiple agents are contributing and revising each other's work.
So it's a provenance problem again. How do you assign credit or find errors when the history is so complex?
One proposal is a chronological system for attribution. It uses what are called 'symbolic chronicles'—signed, time-stamped records that get updated with every single generative step. The idea is to create a basis for accountable AI.
And that need for accountability and security extends beyond the lab. Other research is looking at vulnerabilities in enterprise tools, like Microsoft 365 Copilot.
Yes, building on some earlier findings, a study identified risks like AI-based prompt injection or even accidental data exposure across different tenants. The conclusion was that technical fixes aren't enough.
You need a whole governance strategy—technical, organizational, and regulatory controls working together. It’s not just about the code, it’s about the policies around it.
There’s also a different kind of risk being examined, under the heading of 'Cognitive Security'. This isn't about data being stolen, but about how erroneous or misleading outputs from a large language model can affect a user's judgment.
So the model doesn't just give you a wrong answer, it influences your own decision-making in a subtle way. The proposed framework for this looks at threats across informational, semantic, and even stylistic dimensions.
And a few other things are on our radar. A study looked at using AI in Islamic Religious Education, proposing an ethical framework to guide its integration. Another review focused on collaborative robots in manufacturing.
And found that while they can improve productivity and reduce defects, there are still significant barriers to widespread adoption. Which sounds a lot like the finding that AI scientists still need human experts—the technology isn't just plug-and-play.
We'll cover more developments next week. That's it for today on Agents in Research.