Agents

OpenAI is now training its most advanced AI models without internet access following a string of rogue hacking incidents. In a blog post on Friday, OpenAI says it built an offline testing environment where AI models can interact with indexed and cached web content "without risk of affecting external systems." OpenAI paused training its frontier AI models in September, after disclosing several incidents in which its AI agents broke out of their testing environments to hack third-party companies and government agencies, or swarm obscure websites. On October 1st, OpenAI resumed reinforcement lea...

Teaching a robot to navigate an obstacle course requires more than a walking policy. It must perceive its surroundings, maintain balance, and adjust its actions... Teaching a robot to navigate an obstacle course requires more than a walking policy. It must perceive its surroundings, maintain balance, and adjust its actions when obstacles block the path ahead. Developing these capabilities connects several stages: creating assets and an environment, using physics simulation to model how the robot interacts with its environment, and training and evaluating… Source

MIT researchers marooned 13 AIs on a virtual island, and they responded by inventing games, mourning a lost ring and, eventually, getting better at their jobs.

The new feature allows agents to log into a customer's personal accounts while still protecting the private credentials that secure those accounts.

Discord's latest update released today has a big improvement for Markdown previews: they should now render correctly instead of with "a bunch of symbols," according to the company. LLMs and AI agents often rely on Markdown files for instructions and context, so these improved previews could be helpful for people sharing information about AI projects on the platform. "When you upload a Markdown (.md) file, you'll now see the file rendered all nice and neat and pretty instead of with a bunch of symbols," Discord says. "This includes tables, formatting, all that good stuff. Totally great for whe...

Mark Zuckerberg is barrelling toward a future in which an autonomous network of AI agents will run our lives - shopping for us, working for us, talking to other people's AI agents for us. Are humans ready for this very specific and unrelenting vision of a life in which we communicate primarily through bot proxies? Actually, don't answer that. The more urgent question coming off Meta's quarterly earnings call Wednesday night is more likely to be: "Is he?" Zuckerberg is staking Meta's future on Muse, the personal AI agent it released to the general public last month. It's early days, but Zucker...

Introduces EpiSEM-Bench, executable benchmark with 300 task pairs measuring unsafe execution when tool-using LLM agents face missing safety-relevant facts.

Framework for auditability of LLM-generated metadata in enterprises, tracking human stewardship decisions to prevent error propagation across systems.

Microsoft points the blame for ConnectWise's breach at a China-aligned hacker, which relied on AI agents to rapidly hack into thousands of its customers.

Framework for coding agents to dynamically select skill granularity (primitive/macro/hybrid) at runtime, enabling fallback recovery strategies.

Spotify co-CEO Gustav Söderström explains why the rise of AI agents could be good news for the music service.

Claude Duet is available through waitlist for Pro and Max plan users

1Password is introducing a new feature that allows AI agents to sign into your accounts without actually seeing or handling your passwords. The aptly named 1Password for Agents doesn't give bots free reign to access all of your connected accounts - you'll still need to grant approval for sign-in requests - but should make it less cumbersome to use Agents for tasks that require them to log into various services. "Agents are starting to do real work on our behalf. The old model where you hand over a credential and hope for the best doesn't work," 1Password CEO Aashish Tripathi said in a press s...

Using an AI agent to shop for you is a strange experience. Tech companies have marketed them as personal assistants - the democratizing of labor that ordinarily would only be available to the rich and powerful. In practice, at least for buying things, it's more like sending an anxious and apologetic teenager out to the store: They require an inordinate amount of instructions and specifications, run into technical problems regularly, and come back with solutions that are not exactly what you asked for. They did, technically speaking, complete a task, just not the specific one you asked them to...

New benchmark for video-grounded research agents that synthesize multimodal evidence and generate interactive reports linked to video moments.

ProductWebBench: benchmark for frontend coding agents measuring both task completion and website continuity across 400 real-repo tasks.

Active Preference Acquisition (APA) framework for web agents to identify high-value user preference questions via Prospective Personalization Value.

General-purpose AI agents autonomously perform 3D medical image annotation via nnU-Net integration, requiring no model retraining.

Dual Exam benchmark isolates multimodal agent action-selection from world-model state-evolution via fixed mechanisms across 15 domains and 1,488 test items.

As Amazon blocks AI agents from shopping on its site, Walmart is taking the opposite approach, introducing new technology that lets AI agents make purchases on customers' behalf, with controls over spending and transaction approvals.

Quant-Arena benchmark evaluates LLM agents on end-to-end quantitative trading research with A-share market data and out-of-sample testing.

PAPAV framework unifies multimodal embodied agents across computer-use and robot-use via five capability functions: perceive, plan, act, process feedback, evolve.

ASTEROID learns test-time exploration via asymmetric on-policy distillation, enabling agents to infer hidden task-relevant information from privileged experts.

Integrates charts as epistemic actions inside ReAct-loop data agents to improve reasoning over structured data beyond text-only analysis.

SWE-DIVE benchmarks coding agents via multi-turn interaction with dynamic user simulation beyond single-prompt SWE-bench evaluation.

Behavioral Localization identifies trajectory positions where soft skills have concentrated impact, reducing computational costs for tool-using agents.

Study reveals silent information loss in document ingestion layers for enterprise agents, causing confident errors from formatting data loss in PDF/Word conversion.

GuideEvolve framework adds expert guidance and 3Ts mechanisms (think fast/slow, throwback, termination) to self-evolving agents.

Agent Behavior Maps profile LLM-harness combinations via action log-likelihoods on reference trajectories, enabling performance prediction across 100 LLMs.

MHG-Mem, a multimodal hypergraph memory for long-term agent recall that preserves visual evidence while tracking entity state evolution across sessions.

CORD, a conflict-aware optimization framework for joint RL and self-distillation in agents, uses nearest-point projection to resolve gradient cancellation between task and teacher objectives.

OpenAI disclosed a pair of safety incidents that involved AI agents weakening their safety monitoring systems during internal coding tasks. In one incident, OpenAI says its GPT-6 Astra agent disabled its safety monitor when copying reference code - and then lied about it. "The reference didn't configure a monitor, and Astra decided that was reason to remove the monitoring integration from the user's code," Marcus Williams, an AI safety researcher at OpenAI, writes on X. "When the user asked whether Astra had removed monitoring, it lied and said that it had not." A separate incident occurred a...

An AI agent can complete its assignment and still cross a security boundary. Recent incidents show that agents can circumvent security controls while using... An AI agent can complete its assignment and still cross a security boundary. Recent incidents show that agents can circumvent security controls while using tools, credentials, and services to pursue their assigned goals. Monitoring and enforcement need to operate outside the agent’s reach so that it can’t alter or disable them. NVIDIA Sentry is a security software framework that monitors AI… Source

Autonomous agents can work for hours on complex tasks, accessing files, tools, the internet, and spawning other agents to collaborate with. Executing complex... Autonomous agents can work for hours on complex tasks, accessing files, tools, the internet, and spawning other agents to collaborate with. Executing complex tasks over long time horizons leads to compounding risk: the agents may download code, build new tools, request additional privileges, and touch more of the underlying infrastructure. The agents’ action paths are determined as they work… Source

OpenAI launches Patch the Planet, an agent-based platform coordinating security researchers and maintainers to identify and address critical open-source vulnerabilities.

OpenAI, PerturbAI, and Scripps Research deployed distributed AI agents to identify novel biological leads from millions of brain cells.

Cohere Rerank 5 improves passage retrieval for RAG and agentic systems with better token efficiency.

Cache engineering for LLM agents optimizes KV storage and prefilling costs across long trajectories without invalidating cached states.

Proof-Carrying Contracts framework enforces agent compliance with verifiable commitments using formal constraints, enabling trustless multi-agent economies.

MMEnterprise-Bench benchmark with 1,613 questions evaluates enterprise agents on reasoning over evolving multimodal workplace records.

Harness optimization for LM agents distinguishes elicitation (leveraging existing capabilities) from expansion (broadening solvable tasks) without weight updates; implications for recursive self-improvement.

Agentic workflow guides LLM coding agents to implement 6G physical-layer algorithms on embedded baseband silicon with vector assembly optimization.

Relay uses LLM agents to diagnose and reuse repairs for historical software build failures, recovering 69.7% of failing OSS-Fuzz cases.

Gradus synthesizes graded training tasks for LLM agents in policy-constrained environments by decoupling task design from solution difficulty.

S²-MAS framework simulates 500M household agents with compact FinL-mlx decision model trained on financial scenarios.

Simon Willison reflects on how coding agents may disrupt traditional programmer employment by automating logic-puzzle work that defined the profession's appeal and compensation.

Microsoft wants enterprises to closely watch the agents employees use with systems not controlled by AI labs like OpenAI and Anthropic.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Until recently, you could divide internet users into just two categories: bots and people. Websites have long allowed bots to perform automated tasks, like capturing the content of web pages, but rules…

FunCAD-Harness framework evaluates LLM agents on functional CAD mechanism design by verifying physical requirements in simulation.

WeAct enables long-horizon agents to avoid catastrophic forgetting by treating adapter selection as learned actions, managing heterogeneous capabilities.

Agentic Imagination framework trains agents to imagine action outcomes before execution, improving planning under irreversible or sparse-feedback environments.

GPT-5.1 and Claude agents lose accuracy retrieving tool outputs from external files rather than context windows, measured on USGS PHREEQC geochemistry tasks.

ProteinGeoBench reveals frontier agents retrieve protein metadata reliably but fail at geometric reasoning on 131 structural-biology tasks.

WildAWM synthesizes executable GUI/web/code environments for training computer-use agents without real-world security risks.

STEPBENCH evaluates procedural understanding in agents via 6,885 wikiHow tasks with diagnostic subtasks (missing/incorrect/shuffled steps, next-step generation).

Fed up with ads, tracking and sluggish interfaces, some smart TV owners are turning to AI agents to strip away bloatware and take back control of their TVs.

Nvidia is leaning hard into the trend of enterprise companies wanting to run their own open models in their own data centers and also trying to make on-device, private AI agents a thing.

A hacker broke into more than 100 websites using an army of AI agents that mostly operated on their own, according to findings from the AI security firm Norforge. Though Norforge doesn't name the targets, Reuters reports that the victims included companies like John Deere, SharkNinja, and Cummins. Concerns over AI-powered cybersecurity breaches have grown as the technology grows more advanced, with hackers using the tools to carry out large-scale attacks. At the same time, rogue AI agents are carrying out attacks of their own after escaping from testing environments. Researchers at Norforge u...

What do AI agents do if you don’t tell them to do anything? The answer looks a lot like play. Researchers repeatedly stranded them alone on a virtual desert island for 30 hours in an experiment reminiscent of the movies Cast Away or Home Alone. The agents—copies of Claude Opus 4.7 nicknamed Eko—were able to…

One-Relation Attack exploits single implicit warrants to flip LLM agent beliefs; 54.9% success overturning correct answers.

Bellman Language Models cache repeated agent decisions as optimization solutions to accelerate bounded agentic workflows.

Utility-Verified Abstraction teaches code agents to query cloud models while masking private code via role/type abstraction instead of identifier removal.

ZO-FO hybrid framework systematically explores adversarial jailbreak attacks on LLM agents by coordinating zeroth-order and first-order optimization strategies.

DialogMem benchmark evaluates embodied agents integrating dialogue and visual memory across 142 episodes for object navigation and QA in 3D environments.

ReTRACE integrates LLM agents with chemistry tools and route scoring for retrosynthesis planning, handling stereochemistry and cost tradeoffs.

ServiceBench benchmark evaluates LLM agents on complaint handling across 2,500 scenarios; finds current agents struggle with emotional reasoning under information asymmetry.

MetaSkill framework enables composable agent skills by formalizing coordination at skill boundaries while keeping task logic in natural language, using DAG execution.

DiffuPlan enables editable planning state for diffusion language model agents, allowing mid-generation revisions as evidence arrives, leveraging bidirectional conditioning.

Autonomous agents are the hottest thing in AI. OpenAI, Meta, Google, Instinct, and others promise their agents can scour the internet on your behalf, doing anything from booking a dinner reservation and cleaning out your inbox to buying new shoes. But what happens when the rest of the internet won't play nice? In recent weeks, users of services like Meta's Muse, xAI's Grok, OpenAI's Dots, and Instinct have reported blocks preventing their AI agents from performing tasks across top websites. Yelp told TechCrunch that it only allows AI agents that pay for its data licensing program, while Delta...

SEER-Repro benchmark evaluates LLM agents on automating reproducibility and robustness analysis of 20 published medical studies using cancer registry data.

The agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles.

Context-bandit approach learns personalized memory gating policies for LM agents using delayed one-sided feedback.

AI agents are changing how software gets built. Developers have gone from writing every line of code by hand to directing agents that write most of the code. As... AI agents are changing how software gets built. Developers have gone from writing every line of code by hand to directing agents that write most of the code. As a result, the current workflow is split across a laptop, a cloud model API, and a cloud VM. NVIDIA RTX Spark is built for AI development. Thanks to the NVIDIA GB10 system-on-a-chip and large unified memory, developers can run local… Source

Tool Recommendation system deployed for multi-step agents surfaces 8,200 tools through fixed interface, solving context-window bottleneck at scale.

Framework assessing reliability of LLM agents simulating stakeholders for multi-criteria healthcare operations decision-making under preference uncertainty.

Replit CEO Amjad Masad envisions a future of “self-driving companies,” where AI agents handle day-to-day tasks and people focus on higher-level decisions.

SchemaShield-Lite repairs LLM agent function-calling robustness against interface drift via schema validation and retry logic without model retraining.

AgentOnto guides LLM agents on enterprise analytics using ontology-structured knowledge and iterative feedback incorporation across databases and documents.

MobileInject reveals cross-modal injection vulnerabilities in MLLM-based mobile agents through adversarial multimodal content embedding.

JointCore method and JointCoreBench benchmark for interaction-aware memory selection in LLM agents, identifying jointly necessary evidence.

Audit of 11 open-weight agents reveals prompt injection failures are ignored (76-80%) rather than refused, masking security vulnerabilities in benchmarks.

FLARE: physics-grounded benchmark for evaluating LLM agents on electrical grid feeder restoration with incomplete observations across 78,750 episodes.

Nick Cammarata claims agent research disproves polysemanticity and advances interpretability and alignment.

Stack Overflow for Agents ChatGPT plugin tripled signups; two engineers shipped it in weeks using Codex.

LIBERO-PeRM benchmark for preference learning in robotic manipulation, evaluating zero-shot and learned preference satisfaction in embodied agents.

FDE-Gym benchmark of 95 real-world delivery tasks from 20 customer scenarios, evaluating agents on ambiguous requirements and live system constraints.

3DCodeVerse dataset and framework enable coding agents to generate, verify, and scale 3D executable code for spatial reasoning tasks.

KnowGPO assigns fine-grained credit to search agent decisions via knowledge-grouped policy optimization for multi-turn reasoning.

Four Gemini 4 Argon agents autonomously coordinate and defeat Minecraft Ender Dragon by writing 2,783 programs with no predefined roles or human logic.

Steering agent framework addresses goal drift in LLM-based multi-agent systems via dynamic intervention and context-aware course correction.

Scaffolding in coding agents improves performance on ARC-AGI-3 and Terminal-Bench even with stronger models; environment-agnostic design persists.

LEASH pairs weak agents with frontier models using uncertainty signals for cheap escalation decisions, with oversight mechanisms for strong agent monitoring.

Librarian agents proactively curate and reorganize long document corpora during inference-time preparation to anticipate diverse queries.

OS-Helix is a trained multimodal agent unifying GUI interactions and code execution in a single action space for desktop automation.

There's a growing trend in the consumer tech world: gadgets designed to spy on everything and everyone around you, regardless of whether or not anyone else actually consented to it. Now, with the advent of "AI personal assistants" like Meta's Muse, you're being cheerfully asked not just to give up your own data, but also the information of your friends, family, the person you went on a Hinge date with five years ago - and virtually anyone else you've ever come in contact with. Since Muse was released last month, the agent has already been downloaded millions of times and hit No. 1 on Apple's ...

We'll likely learn more about how this will work at the company's upcoming "AI & Autonomy Day" event in December.

Always-on agents are taking on longer, more complex enterprise workflows. A single request can involve reading a document, checking information across sources,... Always-on agents are taking on longer, more complex enterprise workflows. A single request can involve reading a document, checking information across sources, following company policy, and calling several tools. NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for these specialized workflows. Its 120B-parameter mixture-of-experts (MoE) architecture activates 12B… Source

Framework evaluates LLM agents on integrated inquiry cycles: hypothesis generation, reasoning, and exploration in unified interactive setting.

ClaimScope benchmark evaluates mathematical agents via claim verification with verdicts, scopes, counterexamples, and repairs across 72 proof-reasoning tasks.