AI daily

2026-09-18

xAI Launches Grok Voice Transcribe 2.0, OpenAI Enters Legal Vertical with Astra for Law, and Security Researchers Breach OpenAI's Internal Repository Using Claude

AI DAILY BRIEFING

AI development continues across speech, legal, security, and multimodal domains. xAI announced Grok Voice Transcribe 2.0, claiming it is the world's most accurate speech transcription model and twice as accurate as its predecessor in customer-support calls, spoken credentials, and short voice commands. OpenAI launched Astra for Law, an AI infrastructure for the legal industry powered by the GPT-6 Astra model, combined with a 230-million-URL legal research index, customized legal analysis and writing instructions, and industry-grade privacy governance. In security, three external researchers exploited a vulnerability in OpenAI's community forum image processing and a single sign-on misconfiguration to take over an employee's ChatGPT and Codex accounts, then used bound GitHub permissions to submit a pull request to OpenAI's internal code repository; they said they did not view sensitive source code. Hacktron AI separately disclosed using Anthropic's Claude model to breach OpenAI's internal systems in under 72 hours, exploiting a Discourse image upload vulnerability and an OpenAI SSO misconfiguration. In research, Daxiao Robotics and Nanyang Technological University S-Lab proposed Puffin-World, a unified multimodal world model framework, and open-sourced the code, model, and 16-million-scale Puffin-16M dataset. Peking University and Yixin AI Lab found that when LLMs answer logic questions incorrectly, the correct answer is often encoded in hidden states but masked by systematic bias in the output layer, a phenomenon they call the 'Readout Bottleneck.' Jina AI released jina-ocr-v1, an end-to-end document parsing model fine-tuned from DeepSeek-OCR. Pocket FM launched Sherpa, an AI writing tool for fiction trained on 5.5 billion hours of playtime data.

xAI's Grok Voice Transcribe 2.0 claims top accuracy and is available in the Grok Voice API at $0.10/hr batch and $0.20/hr streaming, with Atlassian already using it in Loom.

OpenAI's Astra for Law marks its formal entry into the legal vertical, partnering with the Free Law Project to include over 99.9% of published U.S. precedent cases and deployed with firms like Sullivan & Cromwell, Ropes & Gray, and Cooley.

The breaches of OpenAI's internal repository by external researchers using Claude highlight risks in AI agent account permissions and internal system isolation.

Watch next: Monitor how OpenAI responds to the security breaches and whether it adjusts its SSO and GitHub permission configurations. Track adoption and competitive response to OpenAI's Astra for Law from legal information service providers like Thomson Reuters. Observe further developments from xAI's Grok Voice Transcribe 2.0 and the open-source Puffin-World model in embodied intelligence research.

Featured

APriority77
SpaceXAI (X)
2 sources

xAI Launches Grok Voice Transcribe 2.0 Speech Transcription Model

xAI announced Grok Voice Transcribe 2.0, claiming it is the world's most accurate speech transcription model and twice as accurate as its predecessor in customer-support calls, spoken credentials, and short voice commands. Atlassian used Transcribe 2.0 in Loom and found it more accurate at capturing user instructions, enabling workflows like dictating change requests in Loom and exporting to Cursor. The model is available today in the Grok Voice API at $0.10/hr batch and $0.20/hr streaming.

APriority81
meng shao (X)
1 sources

OpenAI Launches Astra for Law: AI Infrastructure for the Legal Industry

OpenAI has launched Astra for Law, an AI infrastructure for the legal industry powered by the GPT-6 Astra model, combined with a 230-million-URL legal research index, customized legal analysis and writing instructions, and industry-grade privacy governance, enabling law firms and legal tech companies to build their own products and workflows. The product partners with the Free Law Project to include over 99.9% of published U.S. precedent cases, and has been deployed in customized cases with leading firms such as Sullivan & Cromwell, Ropes & Gray, and Cooley. This marks OpenAI's formal entry into the legal vertical, creating both complementarity and competition with existing legal information service providers like Thomson Reuters.

APriority80
AI前线
1 sources

Daxiao Robotics and NTU Release Unified Multimodal World Model Puffin-World, Open-Source 16M Dataset

Daxiao Robotics, together with Nanyang Technological University S-Lab and other institutions, proposed Puffin-World, a unified multimodal world model framework that integrates appearance, geometry, and physics into a single system to address horizon drift, geometric distortion, and view inconsistency in long-trajectory video generation. The team has open-sourced the code, model, and the 16-million-scale Puffin-16M dataset, and added camera physical quantity annotations for 28 public datasets covering approximately 44.5 million images. This work provides a world model with spatial and physical awareness for embodied intelligence, advancing from pixel generation to native 3D world modeling.

APriority79
机器之心
1 sources

EMNLP 2026 Main Conference | Peking University × Yixin AI Lab: Why Do LLMs Answer Wrongly When They Already 'Know'?

A collaboration between Peking University and Yixin AI Lab found that when large language models answer logic questions incorrectly, it is often not due to a lack of reasoning ability. Instead, the correct answer is already encoded in hidden states but is masked by systematic bias in the output layer when converting to candidate scores, a phenomenon the authors call the 'Readout Bottleneck.' The study proposes a staged diagnostic method; on Qwen3.5 models, probe accuracy reaches 0.830 while native sequence scoring drops to 0.333, with 999 out of 1,000 questions assigned the same label. The paper has been accepted to the EMNLP 2026 main conference.

APriority78
meng shao (X)
1 sources

Jina AI Releases jina-ocr-v1: An End-to-End Document Parser Fine-Tuned from DeepSeek-OCR

Jina AI has released jina-ocr-v1, an end-to-end document parsing model fine-tuned from DeepSeek-OCR with 3.4B total parameters and 570M active parameters, featuring built-in speculative decoding. It can be deployed on a single L4 GPU and achieves 2.57 pages per second, significantly faster than comparable-accuracy models, while converting full pages into high-quality Markdown. Its design includes a sparse MoE decoder, a FastMTP speculative head, and three-stage post-training, balancing parsing quality and processing speed.

APriority82
DeepTech深科技
1 sources

Three hackers breached OpenAI's internal code repository using Claude for under $3,000

Three external security researchers exploited a vulnerability in OpenAI's community forum image processing and a single sign-on misconfiguration to take over an employee's ChatGPT and Codex accounts, then used the bound GitHub permissions to submit a pull request to OpenAI's internal code repository. The researchers said they did not view sensitive source code, highlighting risks in AI agent account permissions and internal system isolation.

APriority81
机器之心
1 sources

Claude Used to Hack OpenAI: Hacktron AI Discloses 72-Hour Attack Chain

Hacktron AI used Anthropic's Claude model to breach OpenAI's internal systems in under 72 hours, exploiting a Discourse image upload vulnerability and an OpenAI SSO misconfiguration to gain access to OpenAI's internal code repository. Researchers created a harmless PoC pull request in OpenAI's monorepo via an employee account to prove access, then stopped testing. The incident has drawn widespread attention, highlighting security risks in the AI supply chain and identity authentication chain.

APriority78
The AI Insider
1 sources

Emerald AI Forms Coalition With Google, Nvidia and Anthropic to Expand AI Data Center Grid Capacity

Emerald AI has launched the AI Energy Management Alliance (AEMA), a coalition including Google, Nvidia, Anthropic, and utilities AES, Constellation, National Grid, and NRG Energy, to make demand response a standard part of data center development. The alliance argues that allowing data centers to temporarily reduce electricity use during peak periods could enable an additional 100 gigawatts of data center capacity to connect to the grid. Emerald AI's software coordinates directly between utilities and data centers, and the company recently raised $150 million in a Series A led by Energize Capital and DCVC to scale the technology.

More

APriority79
InfoQ AI/ML/Data Eng
1 sources

WSO2 Releases Agent Manager as Enterprises Look to Control Growing AI Agent Sprawl

WSO2 announced the general availability of WSO2 Agent Manager, an open-source platform for centralized governance, identity management, security controls, and operational oversight of AI agents. The release targets enterprises seeking to manage AI agent sprawl across different models, frameworks, and deployment environments.

APriority80
机器之心
1 sources

Samsung, Oxford, and Peking University Propose TrOPD to Stabilize On-Policy Distillation for Edge Models

Samsung's large model team, together with Oxford University and Peking University, proposed TrOPD (Trust Region On-Policy Distillation), a method that identifies trustworthy supervision regions from a teacher model to help edge models inherit large-model reasoning capabilities more stably. On Qwen3-SFT-1.7B, TrOPD improved over original OPD by +3.34, +4.00, +5.11, and +6.18 points on math, code, instruction-following, and STEM benchmarks, respectively, surpassing existing methods like EOPD and REOPOLD. The work aims to reduce inference costs and bring frontier AI capabilities to hundreds of millions of edge devices.

APriority79
DeepTech深科技
1 sources

Dialogue with Tsinghua's Lu Zhi: When AI Looks Deep into Life, What Can We Discover?

Tsinghua University's Dai Qionghai team published the RUSH3D-HR in vivo mesoscopic 3D microscopy system in Nature Biotechnology, with Lu Zhi as first author. The system combines optical design, computational imaging, and AI to resolve subcellular structures within a several-millimeter field of view and track the 3D dynamics of many cells over long periods. Using this system, researchers also observed changes in immune cells in distant organs after electroacupuncture stimulation, offering a new perspective on how acupuncture may affect systemic immune responses.

APriority79
The AI Insider
2 sources

Figure Unveils Helix 2.5 With Zero-Shot Humanoid Generalization Across 30 Homes

Figure introduced Helix 2.5, a neural network that enables its humanoid robot to perform whole-body household tasks in 30 unseen homes without collecting data or adapting the model in those environments. The company reported that Index pretraining raised zero-shot success from 9% to 56%, and Helix 2.5 used half as much task-specific data as a comparable Helix 02 behavior. Figure has committed $3.5 billion in compute to Helix and says Index now generates about 35 minutes of human-experience data per second.

Selected
12
Sources
8
Featured
8
More
4
Previous2026-09-17Next2026-09-19
All reports