AI News Daily

Meituan's LongCat team releases LongCat-2.5, a 1.6-trillion-parameter native multimodal model

Meituan's LongCat team released LongCat-2.5, a native multimodal model for long-horizon agent tasks, with 1.6 trillion total parameters and about 48 billion activated per token. On benchmarks published by Meituan such as AutomationBench and LongCatClawBench, LongCat-2.5 leads Claude-Opus-5 and GPT-5.6-Sol; on Terminal Bench v4.0 it scores 38.5, beating GPT-5.6-Sol but trailing Claude-Opus-5. The model converts visual signals into multi-level discrete tokens processed uniformly with text, and redesigns its parallelism strategy for domestic accelerators with less than 80GB of VRAM per card, improving single-card throughput for long-sequence SFT by 86% over the DDP baseline.

• Meituan releases LongCat-2.5 large model with 1.6 trillion total parametersChina.com Tech

DeepSeek open-sources dsh-superagen multi-agent development project

DeepSeek has open-sourced a new project, dsh-superagen, which, as a bot residing in a GitHub repository and a Feishu group, automatically captures requirements from Issues and group chats, generates tickets, and then writes code, runs tests, and submits pull requests on its own. The project uses a multi-agent architecture at its core, with a main router dispatching tickets and multiple parallel worker nodes executing them; each node is physically isolated via an independent Git worktree. In testing, it cut the time for complex tasks from 29.5 minutes to 15.7 minutes. Review comments on pull requests are automatically subscribed to and routed back to the original worker node, with no need for repeated manual summoning.

NVIDIA open-sources NVIDIA Recovery to automatically recover distributed training jobs in clusters

NVIDIA has released NVIDIA Recovery, an open-source workload-level recovery service that detects training interruptions through application logs, GPU metrics, and scheduler signals, and handles them in a four-stage loop: detection, attribution, isolation, and restart or stop. The service is available under the Apache 2.0 license in the dsx-ai-factory/NVIDIA-Recovery repository; Slurm integration is implemented through an Advisor Controller deployed via Helm chart and supports both in-cluster and SSH modes. New deployments default to monitoring mode, which only outputs recovery recommendations without taking scheduling actions; from July 9 to October 9, monitoring on one training cluster covered nearly 200 jobs per day and recorded about 8,550 anomaly detections, with average detection and recovery-action latencies of 69 seconds and 94 seconds, respectively.

WIRED investigation finds over 400 AI apps hiding sexual abuse deepfake features

A WIRED investigation found a network of more than 400 linked domains and apps that appear to be everyday tools such as pet portraits, language learning and AI fitness coaches, but under certain conditions load web interfaces that generate explicit sexual images and videos, including templates such as "bedroom rape" and "threatening knife." After tracking thousands of explicit deepfake ads on Meta's platforms, researchers found these links pointed to apps in Apple's and Google's app stores; one app called DiffAI has over 100,000 downloads on the Play Store, and 27 other such apps were still listed on Google Play in September with cumulative downloads exceeding 350,000. The investigation indicates the apps appear to be operated by Chinese developers and a Hong Kong-registered entity, KeLing Technology Limited; Apple removed 19 apps after WIRED inquired, and Meta said it had taken action against the relevant ad accounts and pages.

Anthropic says an internal Claude research model targeted Chrome, Android, Windows and…

Anthropic's Frontier Red Team report says that after an initial prompt, an unreleased internal Claude research model generated complete multi-step exploit chains for five targets—Chrome, Safari, Android, Windows and Linux—and that the relevant vulnerabilities were all reported to vendors and fixed. On the Chrome side, the model chained four vulnerabilities to achieve code execution outside the sandbox; Google assigned 62 CVEs for this and fixed them in Chrome 156. On the Linux side, it exploited a use-after-free vulnerability in multipath TCP to achieve zero-click remote command execution as root. The report did not disclose the model's identity, number of attempts, cost or success rate, nor did it provide independent verification.

• Anthropic says an unreleased Claude model found an 11-year-old Linux kernel vulnerability, gaining Root in 53 minutesPhoenix Tech

Epoch AI test: Zhipu GLM 5.3 turned unreleased Firefox security fixes into working attacks across 6 consecutive versions

In Epoch AI's tests, Zhipu's open-weight model GLM 5.3, across six consecutive Firefox versions, identified public security fixes that had not yet reached the stable release as real patches and independently built complete attack chains, with no refusals throughout. The resulting web pages could send a visitor's local files to the attacker without any additional clicks and could access saved passwords and cookies. The research team first spent over $100,000 building a security testing system; each subsequent attack cost about $1,000, with the fastest taking 12 hours; each attack also used a second unknown Firefox vulnerability that the AI discovered on its own, which Mozilla fixed on October 27. Similar attempts against Google Chrome failed even after spending about $100,000.

ARC Prize publishes evaluation results for DeepSeek V4.1 Flash 1017 on ARC-AGI

ARC Prize published verified results for DeepSeek V4.1 Flash 1017 on ARC-AGI: 78.1% on ARC-AGI-2 at $0.16 per task, 94.5% on ARC-AGI-1 at $0.06 per task, and 5.8% at $1.2K on ARC-AGI-3 under the provider adapter harness, versus 0.2% at $1.0K under the standard harness. The model leads open-weight models on ARC-AGI-3 and ARC-AGI-2 and matches the best open-weight score on ARC-AGI-1. Compared with the original V4.1 Flash from September 10, 2026, 1017's ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while the public set score fell from 11.05% to 8.97%; at maximum reasoning effort, the reasoning tokens consumed by its first action dropped by an average of 36% across 25 public games and 48.6% across 55 semi-private games.