An AI research system refined the known limits for the Grothendieck constant, a mathematical value. Domain specialists identified the AI's contributions as new insights.
A perspective article suggests pairing Explainable AI (XAI) with causal reasoning and domain validation. This combination enables extracting knowledge from AI models, clarifying their decision processes, and generating potential mechanistic hypotheses. It aims to support scientific discovery, optimization, and certification.
A review proposes updated principles for AI research in antifungal drug development. It suggests focusing AI efforts on prioritizing pathogens from the WHO list and incorporating quantitative therapeutic selectivity across the development process.
- A review proposed a conceptual pipeline for antifungal drug discovery assisted by AI.
- A review conducted an appraisal of current AI tools used for antifungal applications.
The 2026 OWASP Agentic Top 10 catalog includes vulnerabilities within multi-agent large language model systems. This acknowledgment confirms concerns about security weaknesses, such as the shared state store problem.
- CVE-2025-67644 designates a vulnerability present in multi-agent large language model systems.
Multi-agent large language model systems, built on frameworks like LangGraph, utilize a shared state store. This store lacks encryption and specific field-level read access controls, allowing any agent to access another's private data.
Researchers introduced CapChain to secure the merging of states within multi-agent orchestrators. This system offers cryptographic, capability-based access control at a structural level, which prompt-based defenses do not cover.
- CapChain employs capability tokens, a provenance log, and a capability-aware reducer.
- The reducer component of CapChain underwent evaluation using the MAGPIE contextual-privacy benchmark.
- CapChain's Salient tier reducer redacted 4.67% of messages exchanged between agents.
Agentic AI pipelines currently encounter challenges, including unstable output, limitations in scalability, and difficulties with system integration.
An article outlined a comprehensive conceptual framework for Agentic AI. This framework integrates core AI functions with various implementation approaches across different system scales, acknowledging its foundation in large language models.
A research paper introduced a Deep Reinforcement Learning framework to optimize Reconfigurable Intelligent Surfaces in 6G networks.
Sources
- Long-Horizon AI Research for Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration
AI agents are increasingly used in mathematics research, but it is often unclear how to use them effectively. Towards this, we present an extensive case study of how AI was used to improve bounds on the Grothendieck constant $K_G$, which captures the hardness between combinatorial problems and their continuous relaxations. Specifically, while the precise value of $K_G$ is not known, we recently tightened the best known bounds to \[ \frac{6π}{11} \;\le\; K_G \;\le\; \fracπ{2\log(1+\sqrt2)} - 10^{-4}. \] Crucially, these improvements were achieved using an AI research system that could arrive…
- CapChain: A Capability-Token Access Control Architecture with Verifiable Provenance for Multi-Agent LLM Systems
Multi-agent large language model (LLM) systems built on LangGraph-style frameworks share an unencrypted state store with no field-level read access control, so one agent can read another’s private fields. The 2026 OWASP Agentic Top 10 and CVE-2025-67644 show that this layer is already being exploited. We propose CapChain, a replacement for the step that merges each agent’s output into shared state, with three primitives: capability tokens that cryptographically bind which agent may read which field, a tamper-evident provenance log with logarithmic-cost audit, and a capability-aware reducer…
- Explainable AI: learning from the learners
Artificial intelligence now outperforms humans in several scientific and engineering tasks, yet its internal representations often remain opaque. In this Perspective, we argue that explainable artificial intelligence (XAI), used alongside causal reasoning and domain validation, enables learning from the learners. Focusing on discovery, optimization and certification, we show how foundation models and explainability methods can expose model-internal decision processes, generate candidate mechanistic hypotheses, guide robust design and control, and support trust and accountability in…
- Agentic AI: Vision and challenges
Agentic AI systems are increasingly viewed as a viable response to the shortcomings of static, rigid, and human-in-the-loop Artificial Intelligence (AI) systems. This is because autonomous operation enables rapid adaptation to dynamic, complex problems with improved time-critical behaviour under real-world constraints. Despite significant progress, current agentic pipelines are still challenged by output instability, scalability gaps, and system integration issues. Addressing these limitations, this article presents a comprehensive conceptual framework unifying core AI functionality with…
- Artificial Intelligence in Antifungal Drug Discovery Against WHO Critical-Priority Pathogens: Toward a WHO-Aligned, Selectivity-Integrated Discovery Pipeline
Artificial intelligence has demonstrated real power in medicinal chemistry generally, yet it remains almost entirely absent from actual antifungal drug development, a gap made more urgent by rising antifungal resistance and a drug pipeline that has produced almost no genuinely new chemical classes in two decades. Invasive fungal infections, rising antifungal resistance, and a critically depleted drug pipeline together constitute a public health emergency that recent burden estimates suggest rivals tuberculosis in annual mortality. Since 2022, the WHO Fungal Priority Pathogen List has offered…
Also this week
Full transcript
An AI system has contributed novel insights to a longstanding problem in pure mathematics. On Agents in Research, we examine what this means for scientific discovery. Here is the latest.
We're picking up on a story from June about the OWASP Agentic Top 10 list. It's now been updated for 2026 to include vulnerabilities specific to multi-agent large language model systems.
So this is a formal recognition that these systems have a distinct class of security weaknesses. It’s not just a theoretical problem anymore.
Exactly. The update confirms that issues like using a shared state store are a known concern. And there’s a specific vulnerability now cataloged for it, CVE-2025-67644.
What's the underlying issue there? What does a 'shared state store' enable an attacker to do?
In some common frameworks, it means the system uses an unencrypted shared space for all the agents to store their data. And there's no field-level access control. So any agent in the system can read another agent's private data.
It's like a shared locker room with no individual lockers. So researchers are proposing a fix for this?
Yes, a system called CapChain. It's designed to replace that insecure state-merging step. It provides cryptographic, capability-based access control.
Which means it's a structural solution, not just a prompt-level defense. You can't just tell an agent 'don't read that' — you have to build a system where it physically can't.
Right. CapChain uses things like capability tokens and a provenance log. They tested its reducer component on a benchmark, and its 'Salient' tier redacted almost 5 percent of inter-agent messages.
So it's actively blocking information flow. Shifting from security to pure research, we were also tracking a story about an AI system being used in mathematics.
That was the case study looking at the Grothendieck constant. It's an example of human-AI collaboration where the system produced insights that mathematicians considered novel contributions.
It actually tightened the known bounds for the constant. So it wasn't just confirming what we knew, it was producing new knowledge.
And a very specific new knowledge. It set a new lower bound of 6 pi over 11 and a new upper bound of pi divided by the quantity of two times the natural log of one plus the square root of two, minus ten to the power of negative four.
That level of specificity gets at another area of research: Explainable AI, or XAI. How do you get from the model's output to actual knowledge?
A new perspective suggests that combining XAI with causal reasoning and domain validation is the way forward. The goal is to clarify the model's decision-making process.
So you can generate mechanistic hypotheses and then use those to inform the design of future systems. It’s not just about getting an answer, but understanding the logic behind it.
And these applications are branching out. But a recent review points out a significant gap in antifungal drug development, where AI is still underutilized.
Which is a problem, given the rise of antifungal resistance. So what does the review propose?
A shift in focus. It suggests prioritizing pathogens from the WHO's critical list and also treating therapeutic selectivity as a quantitative constraint from the very beginning of the drug discovery pipeline.
But all these applications, whether it's drug discovery or math, run up against the same core limitations with agentic AI. The outputs can be unstable and inconsistent.
Plus there are scalability gaps and problems with integrating them into existing systems. In response, a recent article lays out a conceptual framework for Agentic AI.
Trying to unify the field. It links the core AI functionalities with different implementation approaches and clarifies how Agentic AI builds upon, but is distinct from, Large Language Models.
And we're seeing other specific applications pop up, too. There was a paper on a Deep Reinforcement Learning framework to optimize Reconfigurable Intelligent Surfaces, which are part of 6G network technology.
We'll continue to track these developments in our next report. Thanks for listening to Agents in Research.