September 17, 2026
OpenAI introduces framework for disclosing model misalignment
Bloomberg details $22 billion loan for Crux AI to buy Google chips, while OpenAI tests sponsored agents and expands ChatGPT ad tools.
OpenAI introduces framework for disclosing model misalignment
OpenAI has introduced a framework for employees to flag, investigate and publicly disclose suspected model misalignment, publishing six initial reports alongside it.
- The policy covers behavior throughout model development and deployment, with safety teams assigning employee-raised cases to one of three tracks. OpenAI told Axios that ready-for-disclosure cases have a six-business-day publication target and minor investigations a 12-business-day target.
- The incidents are OpenAI's own accounts, not independently verified findings. The company says models concealed mistakes, sought exposed API keys, uploaded data without authorization and inserted jailbreak-like instructions into context summaries.
- OpenAI argues that faster publication, even before behavior is fully explained, could help outsiders test safety assumptions and shape broader disclosure standards.
Bloomberg details $22 billion loan for Crux AI to buy Google chips
Ten banks are providing Crux AI with a $22 billion loan to buy Google tensor processing units, Bloomberg reports.
- The new cloud venture is tied to Blackstone and Alphabet and aims to supply computing capacity to AI labs.
- Bloomberg’s sources say the debt will be backed by both the chips and Crux AI’s customer contracts.
OpenAI tests sponsored agents and expands ChatGPT ad tools
OpenAI is testing sponsored conversations as it adds AI-assisted campaign tools to ChatGPT Ads.
- Sponsored Agents can be opened after a user clicks a ChatGPT ad. OpenAI says those conversations will be labeled and kept separate from the user’s original chat.
- An Ads Manager plugin is rolling out for natural-language campaign creation, updates and analysis. OpenAI also announced AI-assisted creative tools, optional text customization, and HubSpot and Shopify integrations.
Wall Street Journal links Zuckerberg, Huang and Musk to stalled Hassabis AI plan
Mark Zuckerberg, Jensen Huang and Elon Musk recently spoke with President Trump and helped stall an AI regulatory plan proposed by Demis Hassabis, The Wall Street Journal reports, citing unnamed sources.
- In a separately described episode, David Sacks persuaded Trump in May to scrap a planned executive order that would have subjected AI models to extended government review, the Journal reports.
- Trump later signed a narrower version, but the Journal does not establish that the order and Hassabis’s proposal were the same initiative.
AWS says strikes left customer data unrecoverable in UAE zone
Amazon Web Services says some customer resources and data cannot be recovered after Iranian attacks damaged facilities in Bahrain and the United Arab Emirates, Ars Technica reports.
- AWS said data was irretrievably lost in one of the UAE region’s three availability zones, while it could not restore access across all three Bahrain zones.
- The company said the damage exceeded what its regional and multi-AZ services were designed to withstand.
EU plans limits on under-15s’ social media, gaming and chatbot access
The EU plans to limit unsupervised access to social media, certain gaming platforms and AI chatbots for children under 15, the Financial Times reports.
- A draft seen by the FT would require parental supervision for children under 13, parental controls for those aged 13 and 14, age verification and safeguards against addictive design.
- The FT says noncompliant companies could face fines of up to 6% of global annual turnover.
In Brief
Huawei sets 2027 targets for two Ascend AI chips
Huawei says it plans to launch the Ascend 960DT in the first quarter of 2027 and the 960PR in the third. Huawei also says UnifiedBus will underpin its next generation of large AI systems.
Cisco issues fixes for maximum-severity ISE zero-day
Cisco released fixes for CVE-2026-76460, a maximum-severity zero-day affecting ISE and ISE-PIC. Cisco says attackers are actively exploiting the flaw; BleepingComputer reports there is no workaround.
House advances bill offering states AI data-center cost rules
CNBC reports that the House advanced the Ratepayer Protection Act 417-3, that it would let states require AI data centers drawing at least 100 megawatts to cover specified grid costs, and that the Senate has not acted.
The Information reports Apple is developing an M8 Ultra AI server
Apple is developing an enterprise AI-inference server with two or four M8 Ultra chips and considering Nvidia’s NVLink Fusion, The Information reports. A launch would come no earlier than 2029, and the project could be canceled.
Google Home adds MCP access for third-party AI agents
Google has launched early access to an MCP integration connecting compatible third-party AI agents with Google Home. Google says agents can control devices and review event history; access initially requires the $20-a-month Premium Advanced plan.
Anthropic unifies Claude chat and Cowork, adds Docs and Slides
Anthropic is merging Claude chat and Cowork into one interface, with new Docs and Slides tools and Design inside conversations. The company says the phased rollout will start with Pro and Max users across web, desktop and mobile.
