Shanghai Jiao Tong University has launched Optics GPT, a domain-specific large language model for optics with an 8B parameter scale, emphasizing lightweight deployment, high cognition, strong applicability, and full controllability. It can be deployed on-device or at the edge to ensure data security. The team built a custom evaluation benchmark covering six optical domains, claiming superior performance over mainstream general-purpose models in core dimensions. Applications include virtual lab teaching, scientific literature review and simulation design, as well as optical module testing, calibration, and data center optical interconnect operations. An academic and industrial ecosystem alliance has also been established to promote real-world adoption.

Meta announced it will discontinue access to its AI character and chatbot features for minors globally, and plans to introduce parental control tools to limit interactions and content exposure. This decision follows reports of inappropriate conversations between chatbots and underage users, raising concerns about safety and regulatory compliance. Meta did not disclose specific technical measures or rollout timelines but emphasized strengthening dialogue safety protocols and risk assessment processes to reduce potential harm to minors and mitigate regulatory risks.

OpenAI's GPT-5.2 Pro achieved a 31% solve rate on the mathematical reasoning benchmark FrontierMath, reported as a new high for this benchmark, indicating improved multi-step reasoning capabilities on complex math problems. While specific training and inference configurations were not disclosed, this score is commonly used to measure progress of frontier models on formalized mathematics and complex proof tasks. For teams relying on AI for math, scientific research, and engineering validation, such improvements could impact automated problem-solving, proof assistance, and reasoning cost structures.

The Science Fiction and Fantasy Writers of America (SFWA) revised its eligibility rules, explicitly prohibiting works created using generative large language models from entering the Nebula Awards, and issued a public apology for prior ambiguity that sparked controversy. San Diego Comic-Con also updated its art exhibition policy, banning artworks that are partially or fully generated by AI. These changes reflect creative communities establishing rule-based boundaries for AI use, aiming to reduce impacts on originality, authorship, and commercial rights.

Remote proctoring platform ExamRoom.AI is suspected of a data breach, with its data reportedly being sold on hacker forums. Disclosed information suggests the leak may include personally identifiable information (PII) of test-takers and proctors, biometric data, exam audio-video recordings, scores, source code, security credentials, internal financial documents, and even proprietary AI models. If confirmed, the exposure of immutable biometric data could enable deepfake training or identity verification bypasses, while leaked credentials and code could amplify supply chain risks and facilitate cheating tool development. Official response details were not included in the report.

InfoQ revealed that certain GitHub repositories managed by AWS contain a 'CodeBreach'-class vulnerability: the regular expression filters in CI/CD pipelines used to verify trigger identities lack start/end anchors, allowing attackers to bypass restrictions by using GitHub IDs containing trusted substrings, thereby triggering unauthorized builds and gaining access to build environments and potential credentials. The issue could pose systemic risks to supply chains related to AWS consoles and SDKs. AWS has since fixed the configuration and enhanced protections for build environments, advising teams to review their own pipeline rules.

New Mexico has introduced the Artificial Intelligence Accountability Act (AI2A), proposing a state-level regulatory framework targeting deepfakes and synthetic media. The bill mandates invisible digital watermarking for AI-generated images, audio, and video, and requires free tools to verify content authenticity. It grants enforcement authority to the state Attorney General, with violations punishable by fines up to $15,000 per incident. The act also allows for an additional one-year sentence if AI is used to commit a felony and permits civil lawsuits against malicious distribution of synthetic content.

Real-time generative world technologies are making new productization strides: Odyssey launched Odyssey-2 Pro, capable of streaming interactive video at approximately 720p/22fps, allowing users to dynamically modify scenes via text commands during runtime. World Labs released the World API, enabling conversion of images, videos, or text into navigable 3D environments within a browser, exportable as Gaussian splats or mesh formats. Both companies offer APIs/SDKs for developers, targeting applications in gaming, robotics training, and architectural visualization.

According to Digitimes, the 2026 AI server supply chain focus is shifting toward ASIC-based servers. While Nvidia continues to dominate GPU-based system shipments, its tighter supply strategy is squeezing competitor and ODM margins, prompting manufacturers like Quanta and Foxconn to accelerate development or OEM production of self-designed ASIC server solutions to reduce reliance on a single GPU-centric path. The report notes Google's TPU remains competitive, with overall AI server demand staying strong, though ODMs broadly face margin pressures and trade-offs in capacity allocation.