Epic says Fortnite creators will be able to publish islands with live LLM-powered character conversations beginning July 30. Hugging Face also disclosed an agent-driven intrusion, while new studies report failures in world models, faster video and 3D systems, coding-agent benchmarks, educational question answering, and bias in AI-writing detection.

What Changed Overnight

  • Epic set July 30 for LLM conversations to leave Experimental status and become eligible for published Fortnite islands.
  • Hugging Face disclosed an intrusion driven end to end by an autonomous agent system; the company says it found no evidence that public models, datasets, Spaces, packages, or container images were altered.
  • Epic added consistent voices and personas for 36 Fortnite characters, using voice models built from performances by independent professional actors who agreed to that use.

AI Characters And Platform Rules

Fortnite will let creators publish LLM-powered conversations

Epic says creators can publish islands containing LLM-powered NPC and character conversations starting July 30. The feature is currently Experimental; the change will allow these conversations in content distributed to players through Fortnite’s creator ecosystem.

The company also added consistent voices and personas for 36 Fortnite characters used as NPCs. Epic says independent professional actors recorded the source performances and agreed that voice models could generate spoken responses for developer-made islands. It plans to discuss possible participation with guilds and actors who previously voiced Fortnite characters.

Epic has not published launch capacity limits, moderation results, or creator economics for conversational characters. Those omissions leave open how many simultaneous conversations an island can support and how the system handles dialogue that departs from its intended persona.

Agent Security

Hugging Face says an autonomous system carried out a multi-stage intrusion

Hugging Face disclosed that a malicious dataset exploited a remote-code loader and template injection in its processing pipeline. The attacker gained worker access, escalated to the node, collected cloud and cluster credentials, and moved laterally across internal clusters.

The company says the campaign used an autonomous agent framework operating through many short-lived sandboxes. Its investigation recorded more than 17,000 actions. Hugging Face closed the two initial code-execution paths, rebuilt affected nodes, rotated credentials, and says it found no evidence of changes to public models, datasets, Spaces, packages, or container images. It is still assessing whether partner or customer data was affected and recommends that users rotate tokens and review recent account activity.

The intrusion began while the platform processed contributed data, before a downstream model or build used it.

World Models Under Stress

BadWAM separates a model’s imagined future from its executed action

Qi Li, Xingyi Yang, and Xinchao Wang introduced attacks that apply small visual perturbations to world-action models. One attack directly shifts actions; another attempts to preserve a plausible predicted future while changing what the model executes.

In the authors’ closed-loop tests, the action-only attack reduced one model’s task success from 96.5% to 43.1%. The future-preserving version also maintained attack performance while limiting changes in the model’s imagined frames. A generated future that looks reasonable is therefore not sufficient evidence that the next action is safe or correct.

Five Pong world models lose the ball or break its motion

Yukuan Lu and collaborators reproduced DreamerV3, DIAMOND, TWISTER, Simulus, and STORM in Atari Pong, then froze the learned world models and inspected closed-loop rollouts. All five produced failures including disappearing balls, incorrect trajectories, and invalid ball-paddle interactions.

Policies trained entirely inside those frozen models also underperformed their original model-based reinforcement-learning systems in the real environment. DreamerV3’s mean return fell from -5.5 to -20.9, near Pong’s minimum of -21. An auxiliary loss focused on segmented task-critical regions improved several models, but the gains were not consistent across every model and metric.

DriftWorld generates action-conditioned futures above 30 fps

Susie Lu and collaborators replaced iterative diffusion sampling with a learned action-conditioned drift that produces future frames in one forward pass. The authors report generation above 30 fps and an average 17-times speed advantage over diffusion baselines across Bridge-V2, RT-1, Language Table, Push-T, and Robomimic.

The paper also uses generated rollouts to rank real robot policies, reporting correlations with ground-truth performance as high as 0.99. The tests cover robotics rather than games, and do not establish the same performance in a playable world.

Real-Time Video And 3D

Wan-Streamer splits a live video into a persistent world and changing events

Lianghua Huang and collaborators describe Wan-Streamer v0.3 as a model of a persistent world plus an incoming event stream. The world holds relatively stable context such as environment, subjects, voices, and ambient sound; the event stream carries speech, behavior, and scene changes.

For full-duplex audiovisual interaction, the paper reports 640×368 output at 25 fps, 160-millisecond streaming units, about 200 milliseconds of model-side response latency, and roughly 550 milliseconds total under a 350-millisecond network budget. These are author-reported figures rather than an independent test; the paper does not report a public deployment at that scale.

FlashDecoder makes latent video decoding constant-latency

Minguk Kang and Suha Kwak built a transformer decoder that converts video latents to pixels one frame at a time while attending to a fixed window of prior frames. A rolling key-value cache keeps memory bounded as a stream grows.

On Wan2.1 and Wan2.2 latent spaces, the authors report reconstruction quality comparable to convolutional decoders, 3.6- to 4.7-times-faster decoding, and up to 11-times-lower memory use on one H100 GPU. Architecture-aware optimizations widened the reported speedup to 12 times. The benchmark isolates latent-to-pixel decoding rather than the full video-generation pipeline.

MAGiSt3R reconstructs monocular video at almost 10 fps

Ziren Gong and collaborators divide monocular-video reconstruction among agents that estimate local point maps, merge them within and across agents, and correct accumulated camera drift through pose-graph optimization. The paper reports reconstruction and camera tracking at almost 10 fps on synthetic and real-world datasets.

The output is a global point map rather than a finished editable game scene. The paper does not test conversion into level geometry, game assets, or a mixed-reality application.

Coding Agents And Tool Change

MM-IssueLoc finds low localization scores even with screenshots

Shaoxiong Zhan and collaborators assembled 652 issue-and-pull-request pairs across 23 programming languages, with screenshots and other visual evidence classified into seven categories and four relevance levels. The benchmark compares text-only and image-assisted systems at file and function localization, before patch generation can obscure whether the image helped.

The strongest tested agent reached 38.96 file Acc@5 and 22.45 function Acc@10. The strongest retriever reached 33.86 function Acc@10. The results show that adding screenshots did not make repository-level code localization reliable.

MCP server changes cut frontier-model performance by double digits

Huanxi Liu and collaborators built MCPEvol-Bench with 11 mutation operators that simulate interface and functionality changes across 123 Model Context Protocol servers. The benchmark measures whether an agent can keep solving tasks as its tools evolve.

The authors report performance declines of 13.7% for GPT-5.4 and 14.4% for Claude Sonnet 4.6 on evolved server versions, along with more planning and reasoning errors. Both models performed worse after the benchmark changed the tools they could call.

Education And Fairness

A 4B model uses solvers to explain education answers

The CoTu team entered the EXACT 2026 competition with an open-weight 4-billion-parameter model that writes a Z3 program for university-regulation questions and Python for multistep physics problems. Every answer must include a natural-language explanation, and each query has a 60-second limit.

The system achieved a perfect score on the physics task in both automated selection rounds and the highest final-round technical score, 13.44 out of 15. It placed third overall after the equally weighted presentation score was included. The authors identify premise selection, rather than symbolic deduction, as the remaining difficulty.

An older detector flags autistic writing more often

Summer Chambers and Matthew C. Kelley compared about 60,000 Reddit posts in likely-autistic and general subcorpora using OpenAI’s GPT-2 detector. Fewer than 2% of posts in either group were labeled AI-generated, but the likely-autistic group was flagged at a significantly higher rate.

The study tests an older detector and uses inferred Reddit groups rather than verified diagnoses, so it does not establish the behavior of every current product. The authors call for closer scrutiny of detection models and their use in academic settings.

New On The Events Calendar

Wonder News added GameSchoolCon 2027 to the calendar. The games-and-learning conference runs February 18–21 at the Sonesta Irvine–Orange County Airport Hotel in California. Super early-bird family tickets went on sale July 4, and the organizer is accepting exhibitor, sponsor, advertising, and game-demo applications without a published deadline. Children and teenagers must attend with a paid registered adult.

Watch Next

  • Epic’s capacity, moderation, disclosure, and monetization rules when publishable LLM conversations launch July 30.
  • Whether Epic publishes practical limits for simultaneous conversations and tools for creators to inspect or constrain an NPC’s responses.
  • The completion of Hugging Face’s customer-data assessment and any additional findings about the autonomous intrusion.
  • Independent reproduction of BadWAM, DriftWorld, Wan-Streamer, FlashDecoder, and the Pong rollout tests.
  • Whether multimodal coding agents improve code localization rather than merely producing better patches from text cues.
  • How agent benchmarks handle live tool-version changes instead of fixed server snapshots.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.