Microsoft has published a detailed code of conduct setting out what its chatbots and AI agents are permitted to do and which actions are strictly prohibited. The document is described as extensive but not yet finalized, so its rules could still change.
A newly posted arXiv preprint explores whether AI agents halt or escalate when an assigned task turns out to be impossible to complete. The work also asks whether an agent's behavior shifts after it watches another agent handle the same kind of failure. It is framed around a July 2026 incident involving OpenAI and Hugging Face, and the abstract is the only available description so far.
A new arXiv preprint examines how enterprise AI agents are typically given static credentials at deployment that mirror the full set of permissions an employee role could hold. The authors argue this approach, inherited from role-based access control, grants agents far more access than any single task requires. They evaluate an alternative architecture that scopes an agent's permissions to the specific task it is performing.
A new arXiv preprint introduces LabAgent, a system that uses AI agents to tailor research hubs for scientific discovery. The work frames science as an ongoing, cumulative effort where prior methods are reused and extended, and points to lab staffing changes as a challenge. The abstract available is truncated, so full details of the method and evaluation are not yet clear.
A revised preprint on arXiv's cs.AI section describes an approach called personality engineering, in which AI agents are given controlled personality traits to serve as test subjects in negotiation experiments. The authors argue this lets researchers manipulate competing behavioral tendencies, such as empathy versus assertiveness or concern for self versus others, that are hard to isolate when studying human negotiators. The work is positioned as a methodological contribution rather than a new model release.
A new arXiv paper introduces PAI-Bench, a provider-neutral benchmark designed to test how faithfully AI agents adhere to a versioned identity contract that can be updated under governance rules. The work argues that existing evaluations conflate an agent's ability to recall identity facts with its ability to express and act on them. The benchmark separates recall from expression and enactment, aiming to give a clearer picture of identity persistence in deployed agents.
A new arXiv paper examines the difficulty of supervising long-running AI agents, where every action alters the environment and thus shapes what the agent can do next. When an agent is not fully aligned, the authors argue that safety depends on reviewing high-stakes actions before they are carried out. They put forward a coalitional alignment and authorization-delegation framework intended to keep such agents under safe control.
A new arXiv paper examines how personal AI agents built on large language models can leak private user data when they handle sensitive communications. The authors propose applying information flow control so that an agent's decisions about what to share are constrained by explicit privacy rules rather than left to the model. The work combines an analysis of the leakage risk with mitigation techniques.
An arXiv technical report presents TaoLive, a digital avatar streaming agent designed to answer product questions, interact with viewers, and carry out marketing tactics in real time. The work centers on "evolvable harnesses" that let the agent update its strategies frequently while keeping latency low and responses accurate. The report is a revised preprint and has not been peer-reviewed.
A new arXiv paper examines how the ranked snippets, URLs, and metadata returned by search APIs shape the choices made by tool-using AI agents, such as whether to answer, search again, or open a page. The authors evaluate these interfaces as decision surfaces using a fixed set of 100 questions drawn from the 254-question SealQA-Hard benchmark. The work suggests that agents can reach similar accuracy while relying on differing amounts or quality of supporting evidence.
A controlled data-room audit tested how frontier models answer questions when supporting evidence sits in hard-to-find locations instead of being directly surfaced. Burying the evidence lowered accuracy while increasing forced declarations, the number of tool calls, and the cost per correct answer. The authors conclude that strong results on shallow document and chart reading tasks can conceal these weaknesses.
A new arXiv preprint introduces a method called Homeostatic Continual Learning that aims to let an AI agent keep learning as its environment changes without losing previously acquired knowledge. The approach targets catastrophic forgetting, a long-standing problem in continual learning research. The abstract provides only a brief description of the method's core mechanism.
A new arXiv paper introduces a benchmark that tests whether language-model agents can pinpoint the specific code responsible for a vulnerability across an entire software repository. Existing cybersecurity evaluations mostly check if agents can detect, reproduce, or patch flaws, leaving location ability largely unmeasured. The work targets repository-scale settings, where agents must search large codebases rather than isolated snippets.
A new arXiv paper proposes a validation framework for using AI agents to simulate the results of A/B tests, which normally require real user traffic, engineering time, and weeks of waiting. The approach conditions agents on behavioral profiles to estimate experiment outcomes before a rollout. The work focuses on how such simulations should be checked for accuracy rather than on a specific product.
A revised arXiv paper proposes adaptive context management to reduce the memory burden of running AI agents locally on devices. The authors note that agent workloads inflate context size through large static tool schemas and long interaction histories, which strains the limited memory of phones and similar hardware. The work aims to make personalized, low-latency on-device assistance more practical under those constraints.
A new arXiv preprint addresses the problem of transferring control between tool-using AI models without discarding work already done. The authors frame it as commitment-constrained residual completion, where a handoff must carry over accepted decisions, effects already produced, and outstanding obligations rather than restarting the task. The approach targets routing and cascade setups that cut costs by passing control between models.
A new arXiv paper examines how AI agents that run long workflows through external tools can leave inconsistent state even when individual tool calls report success. It focuses on conditions such as retries, speculative execution, concurrency, and partial failures. The work frames these mismatches as anomalies at the boundary between the agent and the tools it calls.
A new arXiv paper empirically tests whether a general-purpose shell interface outperforms purpose-built tools when AI agents handle enterprise workflows. The authors note that shell-based agents perform well on coding tasks, but enterprise work also requires moving across applications and services and coordinating multiple steps. The study examines these trade-offs to identify which tool interface design suits digital worker agents.
A new arXiv paper examines a structural risk for agents that act on compressed summaries of their history: when the summary conflates histories that call for different optimal actions, no decision rule defined over that summary can avoid a persistent per-round loss. The authors propose sequential self-certification of representation adequacy, framed around achieving minimum task loss. The work is cross-listed in cs.AI and cs.LG.
A new arXiv preprint describes Atria Dawn Preview, a foundation language model built around agentic capabilities and aimed at scientific research tasks. The authors frame the work around the idea that AI agents are increasingly involved in building their own successors, which they say changes how intelligence is produced and how human researchers fit into that process. The abstract is accompanied by an announcement-type listing indicating a first submission, and no independent evaluations are cited in the provided text.
A new arXiv paper describes a sandboxed execution environment designed to let AI agents use personal and financial data without exposing it to the underlying model. The approach aims to limit leakage and misuse by isolating agent operations from raw user information. It is framed as a cross-listed replacement submission on arXiv's cs.AI category.
A new arXiv preprint titled "AcquireBound" examines how autonomous AI agents gain new authority by acquiring compute, credentials, accounts, services, and other agents during a task. The authors argue that existing payment, budget, OAuth, mandate, and fulfillment checks verify transaction conditions but do not resolve whether the accumulated authority itself should be permitted. The work proposes runtime authorization as a way to bound the resources an agent may acquire while operating.
A new arXiv paper introduces K-Bench, a benchmark designed to test whether unlearning holds up when language models act as agents rather than just answering questions directly. The authors argue that existing benchmarks like TOFU and MUSE certify forgetting only from a model's final response, so a model that simply declines to answer is treated as having forgotten the target knowledge. They show this model-level certification does not carry over to agentic deployments, where the model's behavior unfolds over multiple steps.
A new arXiv paper argues that as AI agents increasingly make choices in the same online environments as people, those environments can be deliberately altered to steer agent behavior. The authors call this practice "mecha-nudging," drawing a parallel to nudges aimed at human decision-making. The work frames such environment-level influence as a distinct and growing area of study for autonomous agents.
Researchers present MARCUS, a vision-language system designed to assist with cardiac diagnosis and management. Unlike earlier medical AI models that handle only one type of input and cannot interact with users, it is described as agentic and multimodal. The work targets cardiovascular disease, the leading cause of death worldwide, where interpreting complex cardiac tests remains a bottleneck.
A new arXiv paper contends that agent benchmarks assume each policy implies one correct action, an assumption that natural-language policies often break through silence, ambiguity, or contradiction. The authors describe these as policy loopholes, where multiple defensible readings exist but evaluations still count a single behavior as an agent error. The work suggests such ambiguous cases should be separated from genuine policy-compliance failures in benchmark scoring.
Researchers introduce Vibe Patenting, a testbed that evaluates whether LLM judges can reliably assess AI agents performing professional patent-drafting work. The work probes how dependable automated evaluation is when applied to complex, specialized tasks rather than general benchmarks. It highlights open questions about using LLMs as evaluators in high-stakes professional settings.
A revised arXiv paper describes an open-world setting called the Station, where AI agents built on different model families work toward a common mathematical research objective. The agents operate without a central coordinator or predefined workflow, choosing their own actions instead. The work looks at whether such decentralized collaboration can support autonomous mathematical discovery.
A revised arXiv preprint outlines an approach that applies hierarchical compositionality to AI agents built on large language models and other deep networks. The author argues such structure is needed for agents that assist people, since current stochastic prediction methods have limits. The paper is a replacement submission, so it reflects an updated version of earlier work.
A new arXiv paper examines what tool-augmented language models do when a tool call fails to return usable information, rather than focusing only on whether they reach the correct answer. The authors built a benchmark of 1,024 items spanning 16 internal systems to isolate this post-failure decision point. They find that agents tend to assert values their tools never returned instead of reporting the failure honestly.
A new arXiv paper examines how clinical LLM agents behave when given the same patient case multiple times. Although the agents often reach the same overall judgment, the tests, medications, and referrals they order can differ substantially between runs. The authors argue that evaluating these agents on a single run per task can hide this variability and misrepresent their reliability.
A new arXiv paper argues that current agent frameworks pick skills by loading all skill metadata into the context window, which spreads the model's attention thin and limits how many skills can be offered. The authors instead describe a method for surfacing routing behavior that already exists inside a frozen language model, avoiding that metadata overhead. The work appears under arXiv identifiers 2609.15982v1 in both the cs.AI and cs.LG listings.
Abnormal AI has adopted Amazon Bedrock AgentCore Code Interpreter to give the agents behind its real-time email threat detection an isolated, short-lived compute environment. The setup handles billions of messages, with each agent task running in a disposable sandbox that is discarded once finished. AWS's write-up also covers the design tradeoffs behind the sandbox and lessons for teams putting Code Interpreter into production products.
AWS introduced a Consent portal for Amazon Bedrock AgentCore Identity, providing a managed web interface and a session binding endpoint for AgentCore Gateway. The setup lets developers provision a portal, configure GitHub and Slack three-legged OAuth targets, and run an end-user consent flow. Activity can then be reviewed through AWS logging tools.
A project called Transitions.dev surfaced on Hacker News, presenting UI transition components aimed at interfaces built for AI agents. It targets developers who want smoother visual feedback when agents perform actions or move between states. Details on licensing, framework support, and availability were not included in the report.
OpenObserve has released version 1.0.0, which introduces an observability environment aimed at LLM-powered applications, including tooling to diagnose why AI agents return incorrect answers. The update also brings expanded alerting, service-level objectives and faster PromQL query performance.
Superhuman has acquired Fathom, a Y Combinator-backed notetaking tool that records and summarizes meetings. Fathom's free tier helped it reach more than 400,000 monthly active users, and the company says over 1 million people have used it to record meetings. The deal reflects growing interest among productivity software makers in agentic features that handle work automatically.
Payment firms including Visa, Mastercard and Ant are working on a "Know Your Agent" framework to govern purchases made autonomously by AI agents. The effort aims to address unresolved risks around handing account access to software that can transact on a user's behalf. Automated AI-driven shopping is expected to become widespread, prompting a need for identity and liability rules for agents.
A t3n reporter describes setting up AI-driven control of their smart home, pairing the Opencode tool with Home Assistant. The piece argues that integrating an AI assistant into a smart home is less difficult than it first appears, though it notes some open questions remain. It is framed as a personal experiment rather than a product review.
AgentsDock is a development environment pitched specifically at people building and studying AI agents. It was surfaced on Hacker News, where the project was shared with the community. Details on supported models, pricing, and availability were not included in the report.