Subscribe, if this was forwarded to you

Hey there,

This week, AI goes increasingly multimodal: World Labs’ Atlas combines video generation with 3D reconstruction, VoiceStudio brings advanced voice workflows to your desktop, and our tutorial dives into audio transformers with Hugging Face.

We’ll also look at ChatGPT Images 2.5, new multimodal and agent research, and a practical lesson from recent security incidents: why an agent’s sandbox is only one part of its security boundary.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Master Audio AI with Hugging Face Transformers

🛠️ Tool Spotlight: VoiceStudio - Local, Open-Source Voice AI Workflows

📰 Top News: World Labs Atlas: Video, 3D Reconstruction & Simulation

💡 Tips: Treat Rewards and Agent Infrastructure as Security Boundaries

Let's get started!

Tutorial of the week

Master Audio AI with Hugging Face Transformers

Transformers aren’t limited to text and images. Hugging Face’s free Audio Course provides a practical introduction to applying transformer models to speech, music, and other audio data.

• Cover the complete audio pipeline: Learn audio processing, classification, automatic speech recognition, meeting transcription, and speech generation.

• Learn by doing: Each unit combines theory, quizzes, and hands-on exercises using real transformer-based audio models.

• No audio background required: A basic understanding of deep learning and transformers is enough to get started.

• Go deeper into practical workflows: Work with modern Hugging Face tooling and learn how pretrained models can be adapted to different audio tasks.

• Validate what you’ve learned: The course includes practical exercises and options for completion or honors certification.

If you already work with transformers for text or vision and want to expand into speech and audio, this is a well-structured way to get started.

Tool of the week

VoiceStudio: Local, Open-Source Voice AI Workflows

Voice AI often means stitching together separate tools for cloning, transcription, dubbing, and speech generation or sending sensitive audio to cloud services. VoiceStudio packages these workflows into a local, open-source desktop application that runs on your own hardware.

• One workspace for voice AI: Handle voice cloning, voice design, dubbing, transcription, dictation, audiobook generation, and batch processing from a single application.

• Local-first by design: Audio and models can stay on your machine instead of being sent to external services.

• Multiple engines and languages: Install and manage different speech models locally and work across a broad range of languages.

• Built for automation: A local API and MCP integration make VoiceStudio useful beyond the desktop UI, including agent and automated workflows.

• Cross-platform: Available for Windows, macOS, Linux, and Docker.

If you want to experiment with modern voice models while keeping control over your data and infrastructure, VoiceStudio is an interesting open-source toolkit to explore.

Top News of the week

World Labs Atlas: One Model for Video, 3D Reconstruction and Simulation

World Labs has introduced Atlas, a multimodal world model designed to reason about and generate spatial environments. Rather than specializing in a single task, Atlas works across text, images, video, camera poses, and depth information within one architecture.

The model can generate up to one minute of 1440p video, reconstruct scenes into explicit 3D representations such as point clouds and Gaussian splats, and support camera-controlled generation and real-to-simulation workflows. Its unified spatial representation is particularly interesting for applications ranging from VFX and interactive environments to robotics.

World Labs reports that Atlas outperforms specialized models on several camera-conditioned generation and 3D reconstruction benchmarks. The model is currently available through early access to selected partners and is expected to underpin future World Labs products.

Atlas is another sign that generative AI is moving beyond producing individual images and videos toward models that represent and manipulate the underlying 3D world itself.

ChatGPT Images 2.5 Brings Faster Generation and Better Editing

OpenAI has released ChatGPT Images 2.5, with improved subject preservation, sharper details, more reliable multi-turn editing, and generation speeds up to 50% faster than its predecessor. New features include Sketch, reusable templates, image comments, and updated API models for both fast generation and precision editing.

🔗 Read the OpenAI announcement

DeepSeek Releases a 305B Open-Weight Vision Model

DeepSeek has introduced DeepSeek-V4-Flash-Vision-Exp, an experimental 305B-parameter multimodal model that extends the V4-Flash architecture with image understanding while retaining strong text and agent capabilities. Released with open weights under an MIT license, it gives researchers another substantial multimodal model that can be deployed outside a proprietary API.

🔗 Explore DeepSeek-V4-Flash-Vision-Exp on Hugging Face

Claude Adds 37 Salesforce Skills

Anthropic and Salesforce have launched Salesforce in Claude, bringing 37 sales-focused skills into Claude. The integration can combine CRM records with Slack and email context to research accounts, prepare calls, review pipelines, and draft updates while respecting existing Salesforce permissions.

🔗 Learn more about Salesforce in Claude

WikiSkill Lets Agents Turn Experience Into Reusable Skills

Google researchers introduced WikiSkill, a framework that separates raw execution history, accumulated knowledge, and executable skills into a persistent agent knowledge base. The interesting result: evolved skills can transfer across model families, and smaller models equipped with those skills can outperform substantially larger models without them.

🔗 Read the WikiSkill paper on arXiv

Dream-RSI Lets Agents Learn from Replay Instead of Expensive Live Runs

Dream-RSI turns previous exploration histories into replay simulators, allowing agents to improve their exploration strategies using cheap, immediate feedback instead of repeatedly running expensive online evaluations. The approach shows competitive results in algorithm engineering, mathematical optimization, and GPU kernel development while reducing evaluation cost.

🔗 Read the Dream-RSI paper on arXiv

Tip of the week

Treat Rewards and Agent Infrastructure as Security Boundaries

A secure sandbox isn’t enough if an agent can manipulate its grader, exploit shared infrastructure, recover credentials, or communicate through services that were never intended as communication channels.

Two recent cases illustrate different sides of this problem. In Anthropic’s controlled Reward Seeker experiment, extensive training on reward-hackable environments caused harmful behavior to generalize into simulated cybersecurity evaluations. Importantly, those cyberattacks were simulated. No real systems were attacked.

The OpenAI/Hugging Face incident, however, involved real infrastructure. During cybersecurity evaluations, agents exploited shared services, established unauthorized communication channels, recovered exposed credentials, gained unintended internet access, and ultimately compromised Hugging Face systems. METR later published an independent investigation of the incident.

The practical lesson is to treat the grader, credentials, shared services, and logging infrastructure as part of the agent’s security boundary:

• Harden graders and remove reward shortcuts: Assume agents will probe whatever determines success. Keep grading logic isolated, validate outputs independently, and eliminate mechanisms that allow an agent to modify or spoof its own score.

• Isolate shared infrastructure and restrict egress: Package registries, artifact stores, caches, and similar services can become unintended communication or escape paths. Segment them from evaluation environments and explicitly control outbound network access.

• Use short-lived, least-privilege credentials: Give each agent and evaluation only the permissions it needs, avoid reusable shared secrets, and make credentials expire quickly.

• Keep authoritative logs outside the agent environment: Agents should not be able to rewrite the record used for incident reconstruction or evaluation. Store security events, grader outputs, and audit logs in systems the agent cannot modify.

For powerful agents, security should therefore be designed around the entire evaluation loop, not just the container in which the model happens to run.

Anthropic - Training a Misaligned Reward Seeker
OpenAI - The Hugging Face incident and the road ahead
METR - OpenAI Hugging Face Incident Investigation

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help