Tracked 306 · Curated 15
Cloudflare revealed its internally built AI engineering stack, running on its own products and used by 93% of its R&D organization in the last 30 days. The stack integrates products like Cloudflare Access, AI Gateway, Workers AI, and Durable Objects, significantly boosting developer velocity with a marked increase in merge requests.
Cloudflare held its first Agents Week, announcing a suite of new products designed for the era of AI agents to build the "agentic cloud." The launches span compute, security, and networking, introducing features like Artifacts (Git-compatible versioned storage), Sandboxes GA (persistent, isolated environments for AI agents), and Cloudflare Mesh (secure private networking). These innovations aim to address the challenges of scaling AI agents for developers and businesses.
OpenAI is scaling its Trusted Access for Cyber (TAC) program with the introduction of GPT-5.4-Cyber, a model fine-tuned for defensive cybersecurity use cases. This aims to address the dual-use problem by implementing verified identity and tiered access, reducing friction for defenders while preventing misuse.
Deck.co makes creating website agents incredibly easy by allowing users to define tasks and access them via API. It facilitates back-and-forth communication with sites using structured data, abstracting away the complexities of running a computer and navigating. This functionality even works with sites protected by MFA.
ByteDance researchers introduced PersonaVLM, a framework transforming Multimodal Large Language Models (MLLMs) into personalized assistants with memory, reasoning, and personality alignment. Presented as a CVPR 2026 Highlight, PersonaVLM improves upon baselines by 22.4% and outperforms GPT-4o by 5.2%.
xAI has launched standalone Speech-to-Text (STT) and Text-to-Speech (TTS) APIs, powered by the same infrastructure used for Grok Voice. The STT API offers real-time and batch transcription across 25 languages, priced at $0.10/hour for batch and $0.20/hour for streaming, featuring speaker diarization and word-level timestamps. The TTS API supports 20 languages and five distinct voices with advanced speech tags for expressiveness, priced at $4.20 per million characters.
This work introduces the first benchmark for multi-user LLM agents, addressing the limitation that most AI assistants are designed for a single user. It tests how models handle conflicting interests, privacy constraints, and coordination when serving multiple principals simultaneously.
Noetik is employing AI Transformers to tackle the 95% failure rate in cancer clinical trials. Their TARIO-2 model, trained on extensive tumor spatial transcriptomics data, can predict gene maps from existing H&E assays. This approach aims to match the right patients with existing treatments, rather than discovering new drugs. GSK has partnered with Noetik, signing a $50 million deal.
Codex's new Chronicle feature improves its understanding of context, enabling it to grasp references like "this" or "that" by recognizing on-screen errors, open documents, or past work. This helps Codex learn developer habits, tools, and projects to refine its assistance and workflow integration.
Moonshot AI has released Kimi K2.6 as an open-weight model, built to match GPT-5.4 and Claude Opus 4.6 on coding benchmarks. It features the capability to run up to 300 agents in parallel.
OpenAI has quietly released O1 Preview, which the author describes as the most significant AI advancement since the LLM era's GPT-3.5, accompanied by a crucial chart. This release likely represents a major bet on reasoning and test-time compute, signaling a substantial breakthrough in AI technology.
GitHub's Agentic Workflows integrate AI agents into Actions for tasks like documentation fixes, test generation, and code refactoring. This article details GitHub's security architecture, which assumes agents may leak API keys, spam, or expose secrets. It outlines a three-layer security model designed to mitigate risks from these unpredictable and potentially manipulated systems, contrasting with traditional CI/CD's shared trust domain.
Following last week's preview release of memories in Codex, the feature is expanded with Chronicle. This improvement uses recent screen context, allowing Codex to assist users with their current work without requiring them to restate context.
Software giant Adobe is introducing a new enterprise agent platform to address the growing pressure from AI-native competitors. This move aims to counter the potential disruption of its existing business model by AI technologies.
Moonshot has released its Kimi K2.6 model, an open-weight 1T-parameter MoE model with 32B active parameters, 384 experts, MLA attention, a 256K context window, and native multimodality. The model is touted to achieve open-source SOTA on various benchmarks, showcasing advanced long-horizon execution capabilities. Kimi K2.6 is seen as a competitive alternative to Claude/GPT, particularly for coding and infrastructure tasks.