I. The Pivot to Production-Grade Agents

The most impactful research from October 2025 can be categorized into three interconnected vectors that collectively address the challenges of industrial-scale agent deployment:

• Scaling and Coherence: Novel memory architectures are successfully overcoming the inherent limitations of large language models (LLMs) in long-horizon planning. Techniques like Context-Folding are moving far beyond lossy summarization methods, achieving operational efficiency and accuracy parity with drastically reduced context requirements.² This directly impacts the cost and reliability of agents in domains like Software Engineering (SWE) and Deep Research.
• Democratization and Parity: Breakthrough training methodologies, notably the GOAT framework, are enabling smaller, open-source agents to compete directly with highly resourced proprietary models on complex, goal-oriented tool use.⁴ This is achieved by automating the costly dataset annotation process required for tool proficiency.
• Reliability vs. Risk: Rigorous, high-stakes benchmarking environments, such as AgentArch and STOCKBENCH, provide a quantifiable "reality check" on agent performance. The observed success rates—peaking at only 35.3% for complex enterprise tasks ⁶ - highlight the critical urgency for developing standardized safety, governance, and adversarial resilience protocols, exemplified by the release of the b3 Benchmark.⁸

The quantitative evidence from this research cycle suggests that the immediate deployment of unsupervised, fully autonomous agents in critical enterprise workflows is technically premature. The discrepancy between market enthusiasm and measured technical performance underscores a strategic mandate for Controlled Autonomy. 

2.1 Scaling Long-Horizon LLM Agent via Context-Folding

Technical Problem and Context Drift

LLM agents executing complex, long-horizon tasks, particularly in disciplines such as Software Engineering (SWE) and Deep Research, frequently encounter performance degradation and failure. This issue stems from two related factors: the limited size of the LLM's active context window, and the diminished ability of the model to retrieve relevant information from the middle of an excessively long context, even when using explicitly long-context models.³ Traditional solutions, which rely on simple summarization to compress interaction history, are prone to losing crucial details or suffering from contextual drift, leading to inefficient execution or terminal failure.²

The Context-Folding Mechanism

The paper, Scaling Long-Horizon LLM Agent via Context-Folding, introduces a robust solution through the Context-Folding mechanism. This is a novel, structured approach to memory management designed to maintain task coherence and operational efficiency. Instead of passive summarization, Context-Folding actively compresses the interaction history into a structured, highly relevant active context schema.² This allows the agent to maintain a deep, longitudinal understanding of the task state without requiring a linearly growing context size.

Quantitative Performance and FoldGRPO

The empirical results demonstrate the effectiveness of this architectural choice. The folding agent successfully achieves parity with, or outperforms, standard ReAct baselines on complex long-horizon tasks while utilizing an active context that is 10x smaller.² Crucially, the folding mechanism significantly outperforms models relying on simple summarization-based context management.³ This performance validation establishes that the issue is not merely context length, but the fidelity and structural quality of memory.

The architecture is strengthened by an end-to-end reinforcement learning framework called FoldGRPO. This framework uses specific process rewards to train the agent to perform effective task decomposition and learn the optimal condensation of its memory state.³ The necessity of explicit RL to teach efficient memory state management confirms the complexity of building truly autonomous, long-horizon agents. The superior performance of structured folding over passive summarization establishes the Memory Folding Layer as an essential new component for any production-grade agent architecture facing multi-step, time-intensive problems.

2.2 DeepAgent: End-to-End Deep Reasoning and Tool Discovery (October 2025 Release)

Shift from Sequential Rigidity to Global Strategy

The release of DeepAgent further validates the architectural direction toward sophisticated memory and high-level strategy. DeepAgent is conceptualized as an end-to-end deep reasoning agent that moves decisively away from the rigid, predefined workflow exemplified by the ReAct's sequential "Reason-Act-Observe" cycle.¹⁰ Instead, this agent maintains a global perspective on the entire task, allowing it to perform autonomous thinking, tool discovery, and action execution within a single, coherent reasoning process.

Autonomous Memory Folding and ToolPO

Mirroring the philosophy of context compression for long-term reliability, DeepAgent introduces its own Autonomous Memory Folding mechanism. This system is designed to combat contextual drift and prevent the agent from becoming entrenched in incorrect exploration paths. By compressing the interaction history into a structured, brain-inspired memory schema, the agent is able to "take a breath," reconsider its overall strategy, and proceed more efficiently.¹⁰

This high-level strategic capability is paired with a vast capacity for tool integration. DeepAgent can tackle general tasks by dynamically searching for and utilizing tools from a massive scalable toolset, successfully integrating over 1600+ RapidAPIs.¹⁰ This capability is enabled by the proposed ToolPO mechanism. The convergence of metacognitive control (self-assessment and memory compression) with advanced tool discovery confirms that next-generation agents require unified planning, execution, and memory management capabilities in a single system.

2.3 GOAT: A Training Framework for Goal-Oriented Agent with Tools 

The Annotation Bottleneck and Parity Challenge

A significant hurdle in developing robust, tool-using LLM agents has been the need for massive, high-quality training data, particularly for goal-oriented queries that require decomposing high-level objectives into complex, interdependent API calls.⁴ The traditional reliance on costly human annotation has disproportionately benefited proprietary, closed-source models (such as GPT-4), making it challenging for smaller or open-source models to achieve equivalent tool-use capability.⁴

GOAT's Synthetic Data Generation

The research GOAT: A Training Framework for Goal-Oriented Agent with Tools introduces a crucial development that addresses this challenge. GOAT (Goal-Oriented Agent with Tools) is a novel training framework that automatically generates synthetic datasets of goal-oriented API execution tasks directly from the target API documents, thereby eliminating the need for expensive human annotation.⁵  The technical core of GOAT lies in its systematic methodology. Given a set of API function descriptions, the framework first constructs an API Dependency Graph. This graph meticulously captures all input-output dynamics, identifying how the output of one API function can serve as the input to subsequent, dependent calls. By extracting connected subgraphs from this dependency structure, GOAT synthetically creates complex, multi-step goal-oriented workflows that mimic realistic user queries and generate the necessary training data.⁵

Impact on Tool Use and Democratization

Through extensive experiments on existing goal-oriented benchmarks, open-source models fine-tuned using the GOAT framework achieved state-of-the-art performance.⁴ The framework not only strengthens the reasoning capabilities of open-source agents but also paves the way for their robustness in real-world tool-use scenarios, enabling them to compete effectively with proprietary counterparts.¹¹ Furthermore, the paper introduced a new standard, the GOATBench, for evaluating these goal-oriented tasks.⁴ GOAT represents a major breakthrough in the democratization of agent training. For enterprises, this means the competitive edge is shifting away from the raw power of the underlying foundation model toward the quality and cost-effectiveness of the synthetic fine-tuning methodology applied to proprietary, domain-specific internal APIs. This paradigm shift dramatically reduces the expense and time required for customization and specialized integration.

III. Multi-Agent Orchestration and Interoperability Protocols

As autonomous systems evolve, the ability to coordinate heterogeneous agents becomes essential for tackling large-scale enterprise automation. October’s research highlights the necessary architectural changes required to move past inefficient centralized coordination models and communication silos.

3.1 Anemoi: A Semi-Centralized Multi-agent System (Revised Oct 10, 2025)

Critique of Centralized Architectures

Traditional Multi-Agent Systems (MAS) often rely on a single, powerful LLM acting as a centralized planner. This architectural dependency creates several critical limitations: rigidity in plan execution, reduced scalability due to reliance on a single point of control, and excessive cost and inefficiency resulting from redundant context passing among agents.¹²

Semi-Centralized Design and A2A Communication

The Anemoi framework proposes a Semi-Centralized Architecture to address these issues. This design shifts control by facilitating direct Agent-to-Agent (A2A) Communication. By enabling agents to interact directly, they gain the capacity to monitor the collective progress, assess intermediate results, identify bottlenecks autonomously, and propose adaptive plan refinements in real-time.¹³

This paradigm successfully reduces reliance on a single planner, enabling more scalable execution and minimizing the inefficient prompt concatenation and information loss associated with centralized context management.¹³ The results confirm the effectiveness of this approach: evaluated on the challenging GAIA benchmark, Anemoi achieved 52.73% accuracy using a smaller LLM (GPT-4.1-mini) as the planner, surpassing the strongest open-source baseline (OWL) by +9.09% under identical configurations.¹² This evidence suggests that for generalist MAS, a hybrid, semi-centralized model focusing on efficient A2A dialogue is architecturally superior, validating a clear development path toward an efficient Internet of Agents.

3.2 Co-TAP: Triple Agent Interaction Protocol

The Problem of Information Silos

The ambition of large-scale MAS deployment in enterprises is continuously hampered by the pervasive problem of the "information silo" phenomenon.¹⁴ This occurs because there is a lack of unified communication protocols among heterogeneous agents, leading to high adaptation costs and difficulties in seamlessly integrating agents developed by different teams or running on different foundational models.

Protocol Standardization

The Co-TAP (Triple Agent Protocol) paper proposes a formalized, three-layered agent interaction protocol designed to enforce standardization across the necessary dimensions of MAS operation:

1. Interoperability: Standardizing the communication formats and message structures to ensure agents can understand each other regardless of their underlying models.
2. Interaction and Collaboration: Establishing clear guidelines for dialogue flow, conflict resolution, and collaborative task management.
3. Knowledge Sharing: Defining unified standards for information exchange and context updates.¹⁴

The focus on formalized protocols (mirroring earlier efforts like the Agent Communication Protocol (ACP) ¹⁵) is a critical indicator of industry maturity. For scalable, enterprise-wide adoption, the "plumbing"—the underlying rules of communication—must be standardized. Without such unified protocols, the vision of complex, orchestrated systems like PwC’s AgentOS cannot move from pilot stages to full production.¹⁶

IV. Validation and Domain-Specific Performance: The Reality Check

The true contribution of October 2025 research is its rigor in quantitatively measuring agent performance against real-world, complex requirements. These benchmarks reveal that despite significant architectural gains, critical performance limitations remain, necessitating caution in deployment strategy.

4.1 AgentArch: Enterprise Architecture Validation

Rigorous Enterprise Benchmarking

The AgentArch benchmark provides the most systematic and comprehensive evaluation of Agentic AI systems for enterprise use cases to date. The study evaluated 18 distinct architectural configurations across state-of-the-art LLMs, examining four critical dimensions: orchestration strategy, prompting implementation (ReAct versus function calling), memory architecture, and tool integration.⁶ The goal was to understand how these dimensions interact within complex multi-agent systems performing corporate workflows.⁶

Critical Performance Gaps

The results provide a sobering assessment of current production readiness.

• Finding 1: Complex Task Failure: The highest-scoring models achieved a maximum success rate of only 35.3% on the more complex enterprise tasks, with even simpler tasks reaching a maximum of 70.8%.⁶ This reveals fundamental limitations in the agent's ability to reliably handle multi-step, integrated enterprise workflows involving databases and business processes.
• Finding 2: Reliability Ceiling: The metric for production readiness—the Pass@K score (the rate of success across multiple attempts)—peaked at an extremely low 6.34%.⁷ This figure quantifiably demonstrates that current agent architectures cannot reliably converge to a correct solution when repeated attempts are permitted, indicating profound stability issues for unsupervised deployment.
• Finding 3: No Universal Architecture: The data confirms that No Universal Architecture exists, as models displayed significant, model-specific architectural preferences that varied based on use case complexity. Additionally, multi-agent systems relying on ReAct orchestration demonstrated consistent underperformance across all tested models.⁷

The quantitative evidence of a 35.3% success rate and 6.34% reliability ceiling in controlled enterprise settings serves as a quantified strategic warning. These figures validate the Gartner forecast that over 40% of Agentic AI projects will be canceled by 2027 due to a lack of clear value and guardrails.¹ The data mandates that enterprises immediately integrate mandatory Human-in-the-Loop oversight for all complex agentic workflows to ensure explainability and prevent financial or operational damage.¹⁶

4.2 STOCKBENCH: Dynamic Evaluation in Financial Trading 

Measuring Real-World Financial Risk

Previous financial benchmarks assessed LLM agent knowledge primarily through static question answering.¹⁸ Recognizing that success in finance depends on dynamic, sequential decision-making, the STOCKBENCH paper introduces a contamination-free benchmark evaluating LLM agents in realistic, multi-month stock trading environments.¹⁸

Agents receive daily market signals—including prices, fundamentals, and news—and must make sequential decisions (buy, sell, or hold). Performance is assessed using rigorous financial metrics, including cumulative return, maximum drawdown, and the Sortino ratio.¹⁸

Performance Against Buy-and-Hold

The evaluation of state-of-the-art proprietary (GPT-5, Claude-4) and open-weight models (Qwen3) revealed a crucial finding: the majority of LLM agents struggle to outperform the simple buy-and-hold baseline.¹⁸

This outcome underscores that excelling at static financial knowledge does not translate into successful dynamic trading strategies.¹⁸ The required combination of strategy, risk management, and psychological fortitude in competitive trading is a domain where generalist LLM agents, lacking specialized financial architectures and temporal reasoning capabilities, are demonstrably insufficient. This finding provides the necessary caution for the application of LLM agents in high-stakes capital allocation and high-frequency environments.

4.3 SR-Scientist: Scientific Equation Discovery 

Agentic Approach to Symbolic Regression

In contrast to the financial market where caution is advised, the scientific domain shows dramatic gains. The SR-Scientist: Scientific Equation Discovery With Agentic AI paper focuses on Symbolic Regression—the autonomous discovery of underlying scientific equations from empirical data.²¹

The SR-Scientist framework operates as an autonomous agent, moving beyond the traditional confinement of LLMs to merely proposing hypotheses within pre-defined search algorithms.²¹ The agent is instructed to optimize equations over a long horizon using a dedicated toolset for data analysis and continuous equation evaluation.²² This tool integration includes specialized algorithms, such as the BFGS algorithm, which the agent uses to optimize constant placeholders within the proposed equation skeletons.²²

Significant Outperformance via RL

The approach yielded substantial quantitative success, outperforming baseline methods by an absolute margin of 6% to 35% on datasets covering four distinct science disciplines. The framework also demonstrated enhanced robustness to noise and superior generalization capabilities of the discovered equations.²²

The paper highlights the development of an end-to-end reinforcement learning framework to enhance the agent’s capabilities.²² This demonstrates that for exploratory tasks where the correct solution space is not defined (e.g., scientific discovery), fine-tuning the agent’s iterative interaction loop via RL is a necessary design pattern for achieving high-precision, generalized results.

V. Agentic Safety, Governance, and Adversarial Resilience

The introduction of autonomy significantly increases the attack surface, particularly by exposing the LLM’s internal reasoning and its ability to invoke high-privilege tools. October’s research establishes the first institutional standards for measuring and mitigating these risks.

5.1 The Backbone Breaker Benchmark (b3): Testing Agent Security (October 2025)

Institutionalizing AgentOps Safety

Recognizing the necessity for standardized, measurable security protocols, a collaboration between Check Point, Lakera, and the UK AI Security Institute resulted in the release of the b3 Benchmark (Backbone Breaker Benchmark). This open-source framework is designed to rigorously test the security of the core LLMs—the "backbones"—that power autonomous agents.⁸

Threat Snapshots Methodology

The innovation of b3 lies in its methodology: "threat snapshots." Rather than attempting to simulate the entire, complex, and variable workflow of an AI agent, b3 hones in on micro-tests that delineate the model's behavior at specific, critical moments where vulnerabilities are most likely to be exploited.⁹ These snapshots focus on key decision points immediately preceding sensitive actions, thus efficiently assessing the LLM’s intrinsic resilience.

The benchmark integrates a high-quality dataset of $\mathbf{19,433}$ crowdsourced adversarial attacks collected through the red-teaming simulator game, Gandalf: Agent Breaker.⁸ These attacks target high-risk vectors, including system prompt exfiltration, malicious code injection, phishing link insertion, denial-of-service, and, critically, unauthorized tool calls.⁹

Architectural Implications for Robustness

Initial testing provided valuable architectural intelligence, indicating that LLMs augmented with explicit reasoning capabilities generally exhibit increased security compared to their non-reasoning counterparts. Furthermore, open-weight models are rapidly closing the security gap with closed-source proprietary models.⁹ The b3 benchmark institutionalizes AgentOps Safety, shifting the focus from peripheral input validation to verifying the intrinsic decision-making logic of the LLM at the moment of sensitive execution.

5.2 Adversarial Red-Teaming and Governance Protocols

The development of the b3 benchmark is paralleled by research confirming the evolving complexity of adversarial attacks and the corresponding need for advanced governance systems.

Research into sophisticated threats, such as Genesis: Evolving Attack Strategies for LLM Web Agent Red-Teaming, demonstrates that malicious actors are increasingly using autonomous techniques to generate attacks against agents operating in dynamic web environments.²³ This dynamic threat landscape necessitates protective measures that move beyond static policy enforcement.

This requirement is met by emerging governance models focusing on Uncertainty-Aware, Risk-Adaptive Access Control for agentic systems.²³ Such systems leverage an LLM-Judged TBAC(Tool-Based Access Control) model. For agents that execute transactions or handle sensitive data, traditional access control fails because the agent's context and goal are constantly changing. LLM-Judged TBAC embeds the access decision directly into the agent’s execution loop, requiring the model to assess the real-time risk of the requested action (e.g., a tool call) before authorization.²³ This paradigm aligns enterprise needs for ethical policy embedding (ISO/IEC 42001) and data lineage verification with necessary runtime controls.¹⁶

VI. Strategic Outlook: Implementation Vectors and Competitive Intelligence

The research of October 2025 provides clear, quantitative guidance on strategic investment and necessary architectural adjustments for developers and enterprise leaders moving into the next stage of Agentic AI deployment.

6.1 The Imperative of Controlled Autonomy

The low empirically measured success rate of autonomous agents on complex enterprise tasks (35.3% success, 6.34% reliability) ⁶ mandates a rationalization of high market enthusiasm. Full, unsupervised autonomy is currently a high-risk proposition, validating Gartner’s forecast regarding project cancellations.¹

• Actionable Mandate: Deployment strategy must immediately pivot to favor Controlled Autonomy. Agents should be leveraged primarily as hyper-efficient co-pilots for complex tasks—such as advanced internal research, due diligence, or liquidity optimization (mirroring examples like J.P. Morgan's LOXM 2.0 integration ¹⁶). This approach preserves the crucial Human-in-the-Loop oversight necessary for explainability and error correction, especially where financial or compliance stakes are high.¹⁶
• Governance Structure: Enterprises must establish formal Agent Governance Boards and define clear Standardized Agent Performance KPIs (latency, accuracy, and alignment) to set objective ROI baselines and ensure auditability against the high expectations for increased AI budgets.¹

6.2 Architectural Mandates for Next-Generation AgentOS

Future Agent Operating Systems (AgentOS) must integrate specific architectural components validated by this month's research to address scalability, cost, and reliability.

• Memory Coherence: For all long-horizon Deep Research and Software Engineering tasks, generic summarization methods must be abandoned. Architectures that prioritize structured memory state management, such as Context-Folding or Autonomous Memory Folding, are essential. These methods not only enhance operational coherence but also provide significant cost efficiencies by reducing the required active context size by 10 times.²
• Scalable Orchestration: The AgentArch findings demonstrating consistent underperformance of standard ReAct for Multi-Agent Systems (MAS) ⁷ necessitate a strategic shift. Deployment strategies for MAS must favor architectures designed for high-efficiency A2A collaboration, such as the Semi-Centralized Anemoi framework ¹², or adopt formalized standardization protocols like Co-TAP.¹⁴
• Specialized Fine-Tuning: The GOAT framework offers a clear path to achieving rapid, cost-effective competitive parity in specialized domains. By utilizing GOAT's synthetic data generation capabilities, organizations can customize open-source models for goal-oriented tool use against proprietary internal APIs without the bottleneck of human annotation.⁴ This shifts the focus of competitive advantage from the LLM vendor to the domain-specific fine-tuning methodology.

6.3 Security as a Foundational Layer (Pre-Deployment Validation)

The quantifiable security risks revealed by the b3 benchmark require the integration of mandatory safety processes into the development lifecycle.

• Adoption of b3: Security and development teams must adopt the open-source b3 (Backbone Breaker Benchmark) methodology as a non-negotiable step for pre-deployment validation. This testing must focus specifically on assessing the LLM’s resilience against unauthorized tool calls and prompt exfiltration at defined threat snapshots.⁹
• Dynamic Access Control: Legal, compliance, and engineering teams must enforce governance through dynamic, context-aware policy models. The implementation of LLM-Judged TBAC is necessary to tie data lineage verification and access control directly to the agent’s real-time assessment of risk, effectively mitigating complex vulnerabilities such as memory leakage and self-modification risk.¹⁶

Works cited

1. Agentic AI Trends 2025: Forecasts, ROI Benchmarks & Enterprise Playbook, accessed November 12, 2025, https://usmsystems.com/agentic-ai-trends/
2. [2510.11967] Scaling Long-Horizon LLM Agent via Context-Folding - arXiv, accessed November 12, 2025, https://arxiv.org/abs/2510.11967
3. Scaling Long-Horizon LLM Agent via Context-Folding - ResearchGate, accessed November 12, 2025, https://www.researchgate.net/publication/396499439_Scaling_Long-Horizon_LLM_Agent_via_Context-Folding
4. [2510.12218] GOAT: A Training Framework for Goal-Oriented Agent with Tools - arXiv, accessed November 12, 2025, https://arxiv.org/abs/2510.12218
5. GOAT: A Training Framework for Goal-Oriented Agent with Tools - arXiv, accessed November 12, 2025, https://arxiv.org/html/2510.12218v1
6. (PDF) AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise - ResearchGate, accessed November 12, 2025, https://www.researchgate.net/publication/395526127_AgentArch_A_Comprehensive_Benchmark_to_Evaluate_Agent_Architectures_in_Enterprise
7. Code for AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise - GitHub, accessed November 12, 2025, https://github.com/ServiceNow/AgentArch
8. Open-source b3 framework to benchmark AI agent security unveiled, accessed November 12, 2025, https://securitybrief.com.au/story/open-source-b3-framework-to-benchmark-ai-agent-security-unveiled
9. Lakera Unveils Open-Source Security Benchmark for LLM Backends in AI Agents, accessed November 12, 2025, https://cxotoday.com/news-analysis/lakera-unveils-open-source-security-benchmark-for-llm-backends-in-ai-agents/
10. ️ DeepAgent: A General Reasoning Agent with Scalable Toolsets - GitHub, accessed November 12, 2025, https://github.com/RUC-NLPIR/DeepAgent
11. GOAT: A Training Framework for Goal-Oriented Agent with Tools - ChatPaper, accessed November 12, 2025, https://chatpaper.com/paper/199556
12. Anemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication MCP server from Coral Protocol - arXiv, accessed November 12, 2025, https://arxiv.org/html/2508.17068v3
13. Anemoi: A Semi-Centralized Multi-agent Systems Based on Agent-to-Agent Communication MCP server from Coral Protocol - GitHub, accessed November 12, 2025, https://github.com/Coral-Protocol/Anemoi
14. Co-TAP: Three-Layer Agent Interaction Protocol Technical Report - arXiv, accessed November 12, 2025, https://arxiv.org/html/2510.08263v1
15. AgentOrchestra: Orchestrating Hierarchical Multi-Agent Intelligence with the Tool-Environment-Agent(TEA) Protocol - arXiv, accessed November 12, 2025, https://arxiv.org/html/2506.12508v4
16. AI Agent Trends of 2025: Entering the Agentic Era of Autonomous Intelligence, accessed November 12, 2025, https://genesishumanexperience.com/2025/10/19/ai-agent-trends-of-2025-entering-the-agentic-era-of-autonomous-intelligence/
17. AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise, accessed November 12, 2025, https://arxiv.org/html/2509.10769v1
18. StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets? - arXiv, accessed November 12, 2025, https://arxiv.org/pdf/2510.02209
19. [2510.02209] StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets?, accessed November 12, 2025, https://arxiv.org/abs/2510.02209
20. StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets? | Request PDF - ResearchGate, accessed November 12, 2025, https://www.researchgate.net/publication/396142834_StockBench_Can_LLM_Agents_Trade_Stocks_Profitably_In_Real-world_Markets
21. [2510.11661] SR-Scientist: Scientific Equation Discovery With Agentic AI - arXiv, accessed November 12, 2025, https://arxiv.org/abs/2510.11661
22. SR-Scientist: Scientific Equation Discovery With Agentic AI - arXiv, accessed November 12, 2025, https://arxiv.org/html/2510.11661v1
23. AI Security Research — October 2025 | by Tal Eliyahu - Medium, accessed November 12, 2025, https://taleliyahu.medium.com/ai-security-research-october-2025-8151aca74958

‍