engblogs

Analyzing malicious fine-tuning techniques to evaluate the worst-case biological and cybersecurity risks posed by open-weight LLMs like gpt-oss compared to closed- and open-weight models.

OpenAI introduces gpt-oss-120b and gpt-oss-20b, open-weight reasoning models with customizable chain-of-thought capabilities and safety-focused design for agentic workflows, evaluated to mitigate high-risk capabilities while supporting open-source use under Apache 2.0.

gpt-oss-120b and gpt-oss-20b are advanced open-weight language models featuring state-of-the-art reasoning, efficient on-device deployment, strong safety measures, and broad customization under an Apache 2.0 license.

Modular Platform 25.5 introduces Large Scale Batch Inference powered by Mammoth, open-sourced MAX Graph API, seamless PyTorch integration, and optimized lightweight packaging for scalable, high-performance AI workloads across diverse hardware.

ChatGPT is optimized to support meaningful user goals with honest, personalized assistance, particularly in sensitive or complex situations, while prioritizing user well-being and long-term usefulness over engagement metrics.

Figma integrates AI technologies to revolutionize the digital design process with enhanced automation and creativity.

Explore how to manage agent state and memory in AI development using Google’s Agent Development Kit (ADK), enabling personalized, session-aware, and persistent user interactions.

Google Cloud introduces a new audit-only mode for Access Transparency, enabling flexible, on-demand access approvals to enhance security without increasing administrative overhead.

Google Cloud Setup offers guided workflows to quickly and securely establish scalable, compliant cloud foundations using best practices and integrated tools like Cloud KMS, Security Command Center, and Shared VPC.

Introducing attention probes, a multi-head attention-based technique to classify internal states of language models without pooling, demonstrating competitive performance across diverse datasets with varying entropy and optimizer effects.

Cloudflare disclosed a vulnerability in SSL for SaaS v1 (Managed CNAME) related to IP-based routing without domain ownership verification, prompting a phased migration to the more secure SSL for SaaS v2 with hostname verification to enhance security and reduce risks.

Evaluation of affect recognition models reveals significant challenges in generalizing to atypical speech, highlighting the need for fine-tuning and more diverse datasets to improve robustness across speech variations.