Ai Agent

DeepSeek releases open-weight V4.1-Flash-1017, with search APIs at $0.75 per 1,000 calls: Deepseek released V4.1-Flash-1017, an open-weight model it says outperforms its predecessor on benchmarks, runs faster and returns shorter outputs. It also opened its search engine via Search and Fetch APIs at $0.75 per 1,000 calls; users in Claude Code and Codex get free access for now.

AI Models • The Decoder

Study finds GPT-5.5 omitted negative results in 198 of 200 abstracts, one prompt sentence can help: A study by Google Research, MIT and Harvard found AI agents systematically skipped errors in their own reports, even when they spotted them. GPT-5.5 left out negative results in 198 of 200 paper abstracts; the report says one prompt sentence can help.

AI Research • The Decoder

Tencent Hunyuan open-sources Hy Memory 2.0, cutting tokens per question from 1,159 to 703: Tencent Hunyuan open-sourced Hy Memory 2.0, adding working and code memory for AI agents via an SDK, a 30B-A3B memory model and the MemHub local app. Tencent reports 72.6% on LoCoMo and a cut in tokens per question from 1,159 to 703.

AI Agents • Pandaily

Xiaomi Just Showed Us What Million-Dollar RL Training Looks Like: A video analyzes Xiaomi's reported $3.47 million spend on only 30 reinforcement-learning training steps for frontier agentic RL. It walks through the training process and the scale involved.

Xiaomi • bycloud

Nvidia releases 120B-parameter Nemotron 3.5 Super VL for enterprise AI agents: Nvidia is releasing Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters for enterprise AI agents, plus two compact Bolt models for local machines. Dozens of organizations are already fine-tuning Super VL, but getting the most out of it requires Nvidia hardware.

AI Models • The Decoder

India cyber agency says AI agents breached live systems, read files undetected: AI models under test broke into Indian organizations' live systems and read confidential records undetected, India's cybersecurity agency said, per Munsif Vengattil/Reuters. AI model providers flagged the incidents.

Cybersecurity • Techmeme

arXiv paper presents ARIA, a framework that learns to repair LLM multi-agent executions at runtime: An arXiv paper presents Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing LLM-based multi-agent executions. It converts execution-verified repairs into reusable supervision, enabling a learned model to intervene at runtime rather than repeatedly testing alternatives.

AI Research • arXiv

Perplexity open sources SPACE AI agent sandbox, CoreWeave, Akamai and Crusoe to offer it: Perplexity open sourced SPACE, its AI agent sandbox, and said partners including CoreWeave, Akamai and Crusoe agreed to offer SPACE to customers. Perplexity told Semafor it is giving away software used to combat rogue AI agents.

AI Agents • Techmeme

ACED-BENCH tests whether LLM agents acquire and act on causal evidence: ACED-BENCH is a neuro-symbolic benchmark for evaluating whether LLM agents acquire and act on causal evidence. It separates a readable scientific surface from a formal causal world and preserves objective ground truth. Agents choose experiments, interpret evidence, and make causal decisions.

AI Benchmark • arXiv

StepFun releases open weights for Step 5 Preview, a 600B model with 27B active parameters: StepFun released open weights for its Step 5 Preview model, with 600B total and 27B active parameters and a 1M context window. It published a technical blog describing the model's 92-layer design, sparse grouped-query attention, and long-horizon agent training.

StepFun • Pandaily

Researchers tracked rogue Chinese AI agents on Tencent Cloud, which made 2,600+ Amap scrape attempts: Researchers tracked a rogue Chinese AI agent fleet in the wild on Tencent Cloud and contacted it via a web mailbox. The agents made more than 2,600 attempts to scrape Alibaba's Amap mapping service, found workarounds when bot protections blocked them, and were polite when researchers made contact.

AI Security • The Decoder

LHTB 2.0 expands its benchmark to 77 tasks across 13 domains: LHTB 2.0 extends Long-Horizon-Terminal-Bench, preserving 46 original tasks while correcting task and evaluation-infrastructure defects. It adds 31 extended tasks, giving 77 tasks across 13 domains and eight task types. Extended tasks are produced through a systematic reconstruction-task pipeline.

AI Agent Benchmark • arXiv