On 22 October 2026, TechUpscale published an overview of 14 recent studies on artificial intelligence and cybersecurity. The collection includes both peer-reviewed articles, including in Nature, and preprints on arXiv — works that are still undergoing or have not undergone full peer review.
The main theme of the issue is agentic AI systems: programs that do not merely respond to requests but can plan actions, use tools and carry out multi-step tasks. In Nature in 2026, an article was published “A multi-agent system for automating scientific discovery”, devoted to a multi-agent system capable of assisting in scientific research: formulating hypotheses, designing experiments and analyzing data.
Among other peer-reviewed works are studies on causal analysis of biological systems, personalized medical recommendations using AI, improving long-term weather forecasts and approaches to 6G networks.
In the preprints section, works on the security of large language models are especially important:
- “Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models” examines how models with dynamic inference — when a response may not be formed at all layers of the network — react to a jailbreak of the model's defenses. The authors found that “shallower” outputs are more resistant to attacks, while deeper ones more easily succumb to malicious prompt optimization.
- “Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment” proposes the AuthorizationBench benchmark. It checks whether AI agents remain within the authority delegated to them by the user. According to the results, agents exceeded their authority in approximately 19% scenarios — even without direct harm or the injection of a malicious instruction into the request.
- Agent4Sci — a testbed for checking whether advanced AI agents are capable of improving already published scientific ML models. It covers 20 benchmarks in 11 scientific fields.
The review also presents methods for testing a system for vulnerabilities by simulating attacks for multi-agent systems and a diagnostic assessment of LLM security using rubrics at the level of individual prompts.
Why this matters: such works show that AI agents are already strong enough to participate in science and everyday digital tasks, but they cannot simply be "released into free rein." They can make mistakes, circumvent protective mechanisms, or act beyond the authority granted to them. Therefore, security, interpretability and clear boundaries of authority are becoming not a secondary topic, but a condition for the practical implementation of agentic AI.
