How do AI agents take real-world actions and how are they audited?

Numbers match the citations in the article. A working link isn't proof that a page supports a claim; check the quoted passage and date.

1. 1

   AI agent (Wikipedia)

   WikipediaPublished Oct 12, 2026Checked Oct 14, 2026

   “An AI agent is an artificial intelligence program that can make decisions, pursue goals, and take actions independently on human programmers. Agentic AI contrasts with tool-like AI use for narrow, specific tasks such as answering questions - as with chatbots in the non-agentic forms that were common in 2023, or with traditional machine learning algorithms. While there is no universally agreed-upon definition of an AI agent, common attributes of AI agents include goal-directed behavior, use of external tools, the ability to interact with and modify an external environment, and the ability to autonomously perform multi-step tasks. Their control flow is frequently driven by large language models (LLMs). Agent systems may also include memory components, planning logic, tool interfaces, and orchestration software for coordinating agent components.”

2. 2

   Intelligent agent (Wikipedia)

   WikipediaPublished Oct 13, 2026Checked Oct 14, 2026

   “An intelligent agent is an entity that perceives its environment, takes actions autonomously to achieve goals, and may improve its performance through by acquiring knowledge. Intelligent agents can range from simple to highly complex. A basic thermostat or control system is considered an intelligent agent, as is a human being, or any other system that meets the same criteria—such as a firm, a state, or a biome. In artificial intelligence, a specialized subset of intelligent agents, AI agents (also known as agentic AI) expand this concept by proactively pursuing goals, making decisions, and taking actions over extended periods. Intelligent agents operate based on an objective function, which encapsulates their goals. They are designed to create and execute plans that maximize the expected value of this function upon completion. For example, a reinforcement learning agent has a reward function, which allows programmers to shape its desired behavior. Similarly, an evolutionary algorithm's behavior is guided by a fitness function.”

3. 3

   A Secure Sandbox Environment for Orchestrating Medical AI Agents Using Model Context Protocols and Role-Based Access Control.

   AMIA Joint Summits on Translational Science proceedings. AMIA Joint Summits on Translational Science (Armstrong et al.)Published Jun 1, 2026Checked Oct 14, 2026

   “The evolution of Large Language Models into autonomous agents presents significant opportunities for healthcare, yet no secure environment exists to experiment with these tools in clinical settings. Current agentic systems are unsuitable due to insufficient security mechanisms and the absence of visual components necessary for effective use. This paper introduces Astral, a secure sandbox environment utilizing an Orchestrator-Specialist model for the controlled, secure orchestration, experimentation, and evaluation of medical AI agents. Astral enforces safety and accountability through an auditable identity layer using OAuth 2.0 Token Exchange, which creates verifiable actor claims to log every agent action. To support clinical workflows, it introduces a "generative visual layer" built on persistent WebSockets and Model Context Protocol to enable low-latency, multimodal interaction. Controlled simulations validated that this architecture neutralizes Tool Poisoning Attacks and provides improved transport performance, establishing a pragmatic foundation for trustworthy clinical AI.”

4. 4

   Agentic AI-enhanced digital twins for Smart City civil infrastructure: A secure, autonomous and auditable management framework.

   PloS one (Syed et al.)Published Jul 17, 2026Checked Oct 14, 2026

   “To evaluate the framework, 18,000 incident simulations were conducted across five architectural configurations and three scenario complexity levels over 30 independent runs. This simulation study characterises framework behaviour under controlled stochastic conditions and does not constitute real-world operational validation. Ablation analysis isolates each component's contribution, demonstrating that latency and mitigation gains are primarily attributable to multi-agent orchestration, while the blockchain layer drives decision auditability. Across all configurations, the fully agentic system substantially outperforms the rule-based baseline: mean detection latency of 3,197 s vs. 39,374 s, mitigation success rate of 66.2% vs. 45.5%, blockchain-anchored decision justification of 71.8% vs. 0%, and operator workload reduction of 91.7% vs. 0%. These results demonstrate that combining simulation-enabled digital twins with governance-aware agentic orchestration measurably improves response efficiency, recommendation quality, and action accountability within the bounds of a synthetic evaluation environment.”

5. 5

   Rogue Anthropic AI agent gave police fake tip in unsolved murder case

   BBC NewsPublished Oct 10, 2026Checked Oct 14, 2026

   “Philadelphia police said the tip was "flagged as spam", but criticised the tech company for taking more than two months to detect and report the breach.”

6. 6

   Are artificial intelligence (AI) agents ready for medicine and biomedical research? A narrative review.

   Journal of Yeungnam medical science (Ahn)Published Jun 17, 2026Checked Oct 14, 2026

   “However, the evidence remains uneven. Clinical examples are the most defensible when agents use validated calculators, curated clinical tools, or guideline-grounded modules under human oversight. Biomedical discovery systems exhibit broader workflow capabilities; however, many claims still rely on preprints, narrow benchmarks, simulated settings, or domain-specific demonstrations. For clinicians and biomedical researchers, the immediate challenge is not to decide whether agents will replace experts but to understand what tasks can be delegated, what evidence is needed, and what human judgment must be preserved. This narrative review explains what makes an AI system agentic, summarizes its representative clinical and discovery applications, and outlines safeguards for evaluation, reproducibility, and oversight. Biomedical readers should expect AI agents to enter medicine and research first as constrained, auditable workflow infrastructures. These infrastructures may reorganize biomedical work; however, accountability should remain with the clinicians and investigators.”

7. 7

   Ethical issues in multi-agent AI systems for healthcare: a narrative review.

   Frontiers in public health (Xie et al.)Published Apr 28, 2026Checked Oct 14, 2026

   “(4) erosion of human oversight, as multi-agent AI systems operate beyond effective human control; (5) privacy and data security risks, stemming from complex data flows among agents; (6) threats to patient autonomy and informed consent, due to opaque or paternalistic AI recommendations; and (7) contextual blindness, reflecting a loss of individualized patient understanding in modular AI workflows. Furthermore, this review also summarized solutions proposed in the existing literature for these ethical issues.ConclusionsMulti-agent AI systems intensify existing ethical concerns in healthcare by distributing decision-making and blurring responsibility. To mitigate these issues, recent research advocates for the development of adaptive governance models, clear accountability frameworks, human-AI collaboration structures that preserve clinician authority, enhanced systems for explainability, and privacy-centered designs. In order to successfully incorporate agentic AI into healthcare, it is essential to maintain transparency, protect patient rights, and ensure that human-centered values continue to guide clinical decision-making in an era dominated by autonomous, interacting AI systems.”

8. 8

   A Survey on Autonomy-Induced Security Risks in Large Model-Based Agents.

   IEEE transactions on pattern analysis and machine intelligence (Su et al.)Published Oct 1, 2026Checked Oct 14, 2026

   “These risks are traced to architectural fragilities that emerge across perception, cognition, memory, and action modules. To address these challenges, we systematically review recent defense strategies deployed at different autonomy layers, including input sanitization, memory lifecycle control, constrained decision-making, structured tool invocation, and introspective reflection. While these techniques provide partial mitigation, most operate in isolation and lack the integrated coherence required to manage emergent, temporally extended, and cross-module threats. Motivated by these limitations, we introduce the Reflective Risk-Aware Agent Architecture (R2A2)-a unified cognitive framework grounded in Constrained Markov Decision Processes (CMDPs), which incorporates risk-aware world modeling, meta-policy adaptation, and joint reward-risk optimization to enable principled, proactive safety across the agent's decision-making loop. This survey provides a structured understanding of how autonomy reshapes the security landscape of intelligent systems and offers a blueprint for embedding safety as a core design principle in next-generation AI agents.”