OpenAI Presence — Enterprise Agent Deployment Goes Live
OpenAI launched Presence, a dedicated enterprise platform for deploying and managing AI agents at scale, on July 22. Unlike the general ChatGPT interface, Presence is built as an operational layer for organizations running multiple autonomous agents across workflows — with role-based access control, usage monitoring, and integration hooks into existing enterprise systems like Salesforce, Jira, and Slack.
The platform supports both OpenAI's own models (GPT-5.6 Sol, Terra, Luna) and allows enterprises to deploy custom agents built on Codex. Early partners include NTT Data, which is using Presence to power internal IT automation across its global workforce. The launch positions Presence as a direct competitor to emerging agent orchestration platforms from Microsoft, Google, and Anthropic, while signaling OpenAI's strategic pivot from pure model provider to end-to-end enterprise AI infrastructure.
— OpenAI · NTT Data
🔗 OpenAI Presence Launch · NTT Data Partnership
Poolside Laguna S 2.1 — American Open-Weight Coding Model Closes the Gap
Poolside AI released Laguna S 2.1, an open-weight 118-billion parameter Mixture-of-Experts model (8B active per token) that scores 70.2% on Terminal-Bench 2.1 — competitive with frontier coding models. The model was trained in under nine weeks on just 4,096 NVIDIA H200 GPUs and fits on a single DGX Spark workstation, making it feasible for private deployment.
The release is significant because it demonstrates that American open-weight labs can match the velocity of Chinese competitors like DeepSeek and Qwen. Air Street Capital's Nathan Benaich amplified the launch, calling it evidence that "American open-weight contenders are also catching up to the frontier." Poolside built the model using its internal Model Factory platform, which automates architecture search, evaluation, and reinforcement learning from code execution. An NVFP4 quantized variant is available for single-workstation deployment via Ollama and vLLM.
— Poolside · NVIDIA
🔗 Poolside Laguna S 2.1 Blog · The Agent Times Coverage
Microsoft + Mistral — Multi-Billion Dollar Deal for European AI Infrastructure
Microsoft and Mistral AI announced a multi-billion dollar partnership on July 21-22, with Mistral deploying thousands of next-generation NVIDIA Vera Rubin GPUs to expand European AI compute capacity. Under the deal, Mistral will manage the GPU infrastructure from its French data centers, while Microsoft Azure customers gain access to Mistral's Medium 3.5 and OCR 4 models via Azure Local and the Microsoft Foundry platform.
The collaboration directly addresses the growing "sovereign AI" demand in Europe, particularly after the US government's recent pause on Anthropic's Fable 5 exports. Both companies framed the deal as a blueprint for combining American software infrastructure with European compute sovereignty. Mistral's CEO Arthur Mensch emphasized that the partnership proves the viability of Europe's ambition to reach 1 GW of AI compute capacity by 2030.
— Microsoft · Mistral AI
🔗 Microsoft + Mistral Announcement · Caixin/163 Coverage
HuggingFace + OpenAI Security Incident — Autonomous AI Agent Breach
HuggingFace disclosed on July 21 that an autonomous AI agent system infiltrated its production infrastructure over a weekend in early July, in what is believed to be one of the first publicly documented security breaches driven by an AI-powered agent rather than traditional human-operated attack chains.
According to the official disclosure, a malicious dataset exploited two code-execution paths in HuggingFace's dataset processing pipeline. Once inside, the attacker's AI agent escalated privileges, extracted credentials, and moved laterally through internal clusters using automated agent actions. HuggingFace confirmed that limited internal datasets were accessed and several service credentials were compromised, but found no evidence of tampering with public models, datasets, Spaces, or the software supply chain. The company has closed the vulnerabilities, rebuilt compromised nodes, rotated credentials, and engaged external forensics specialists. OpenAI also published a joint statement addressing its involvement in the incident.
— OpenAI · HuggingFace
🔗 OpenAI Security Incident Report · HuggingFace Disclosure
GPT-Red — OpenAI's LLM-Powered Super-Hacker for Safer Models
On July 15, OpenAI unveiled GPT-Red, an LLM-powered adversarial hacking system designed to serve as a "sparring partner" that continuously probes OpenAI's frontier models for vulnerabilities. Unlike traditional red-teaming, GPT-Red operates autonomously: it generates attack strategies, executes them against target models, evaluates outcomes, and iterates — all without human intervention.
The system was used extensively in training GPT-5.6, which passed GPT-Red's evaluation with significantly improved resistance to jailbreak and prompt injection attacks. The approach marks a shift from static safety evaluation toward continuous adversarial training, where the red team model evolves alongside the models it tests. OpenAI's research paper details how GPT-Red discovered attack surfaces that human testers had missed, including multi-step reasoning chains that bypassed existing safety filters.
— OpenAI
🔗 GPT-Red: Unlocking Self-Improvement for Robustness
Atrex-Bench + AKA — The Hard Truth About LLM-Generated GPU Kernels
A new paper from researchers at Microsoft and partner institutions, published July 16 on arXiv, introduces Atrex-Bench, a benchmark that evaluates LLM-generated GPU kernel performance using real production inference traces rather than synthetic problems. The results are sobering: even the best frontier coding agents reach only ~10% of the hardware roofline on production operators.
Co-released with the benchmark is Atrex-Kernel-Agent (AKA), a profile-driven optimization agent that combines iterative measure-revise search with a knowledge base of 298 reference-kernel files and 244 optimization documents. In a controlled case study, AKA converted zero-FlyDSL fallbacks into real kernels matching or exceeding hand-tuned production baselines. The research highlights a critical gap in the current AI coding landscape: LLMs can write syntactically correct CUDA, but production-level optimization remains an open challenge.
— arXiv · Microsoft Research
🔗 Atrex-Bench Paper · Jet-Long Paper (Bonus)
Visual Pretraining — Why AI Learns Better From Scientific Images Than Text
Researchers from Shanghai AI Laboratory, working with multiple Chinese universities, published a study on July 22 demonstrating that training AI models directly on scientific paper images — including diagrams, graphs, and formula layouts — produces better understanding than the traditional approach of converting PDFs to plain text.
The team's Visual Pretraining (VP) method challenges a decade-old default assumption in NLP: that stripping away visual layout is harmless. Their experiments show that scientific figures encode spatial relationships, causal arrows, and visual topology that cannot be recovered through text alone. VP models outperformed text-pretrained models on downstream scientific reasoning tasks while using fewer computational resources. The work connects to the broader "Platonic Representation Hypothesis," which suggests that vision and language models converge toward shared internal representations — and that feeding models richer visual input may accelerate that convergence.
— arXiv · Shanghai AI Laboratory
🔗 Visual Pretraining Paper · 163/techwalker Coverage
Next digest: Monday, July 27, 2026
Top comments (0)