Latest / HotTea: The AI Morning Briefing / The gaps in OpenAI's agent monitoring
Transcript
- Sections 00:00 What we're covering today 00:28 1. Outside testers found ways around OpenAI's agent monitoring. The fixes remain unverified 01:31 2. Commerce proposes $389 million for power chips and materials, with equity stakes attached 02:28 3. Four companies move agent projects into the Agentic AI Foundation 03:18 4. Claude's new Google Cloud route still makes Anthropic the data processor 04:14 5. OpenAI opens a faster API for questions with a fixed set of answers 05:07 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Hot Tea for Friday, October twenty-third, twenty twenty-six, narrated with an A I-generated voice. On Thursday, October twenty-second, Commerce proposed research awards for power chips, with government equity attached. Four companies have put agent projects under shared governance. Anthropic's Google Cloud service has separate data-processing terms. And OpenAI is promising faster answers when developers define the choices in advance. Outside testers found ways around OpenAI's agent monitoring. The fixes remain unverified. METR published its October 22 review of OpenAI's internal agent monitoring. One METR researcher and one Redwood Research researcher spent three weeks onsite, from the twenty-first of July through the eleventh of August. They found ways agents could, in principle, evade the system, including an exploit that could hide hundreds of recent actions. The review assessed the system as it existed on and before the twenty-first of July, while OpenAI was expanding it. So it does not establish the current system's coverage. OpenAI says it patched every identified vulnerability affecting production safeguards, but METR has not evaluated that claim. The public summary is limited, and OpenAI could redact sensitive material. These were tests of possible bypasses, not proof that agents used each one. The next useful check would independently test the fixes and measure monitoring coverage. A plan to expand coverage does not establish that every intended action was reviewed. Commerce proposes $389 million for power chips and materials, with equity stakes attached. The U.S. Commerce Department announced four letters of intent on October 22. Together, the proposed awards could provide up to $389 million for semiconductor research. Vertical Semiconductor could receive up to $73 million for vertical gallium nitride power devices. Its proposal includes a 2.5-megawatt A I data center rack demonstration. These are still proposed awards under letters of intent. If they are finalized, Commerce would receive a minority, non-controlling, non-voting stake in each company. The announcement funds research. It does not show that the devices meet their targets or that the rack is operating. We'll learn more from final award terms and demonstrated power-conversion performance. The other proposed recipients are Helion, Commonwealth Fusion Systems, and Sibelco. Four companies move agent projects into the Agentic A I Foundation. On October 22, the foundation announced four new projects. Anthropic contributed Agent Skills, OpenAI contributed Open Responses, Hugging Face contributed OpenEnv, and Alibaba contributed OpenSandbox. Together, the projects cover reusable instructions, model interfaces, training environments, and sandbox execution. Shared governance gives more developers a role in these projects. But it does not make every agent product compatible, and it does not show that a sandbox is secure. Teams still need to check how each implementation handles instructions, tools, and permissions. Agent Skills packages instructions, scripts, and reference material that an agent loads when needed. Specifications and tests across products will show whether implementations work together. Claude's new Google Cloud route still makes Anthropic the data processor. Anthropic says its managed Claude service on Google Cloud became generally available on October 22. Customers use Google Cloud access controls and billing, and usage can count toward their cloud spending commitments. Anthropic says it operates the service and acts as the data processor. These are Anthropic's service claims, not independent evidence of customer results. Google's managed option is separate for customers who need Google as the data processor. Anthropic's documentation lists unsupported features, and it says a zero-data-retention agreement from another Anthropic organization does not carry over. A shared cloud bill does not make the services equivalent. Before moving a workload, check the processor, retention arrangement, supported features, and location requirements. The documentation currently lists United States and Global endpoints for the Anthropic-managed service. OpenAI opens a faster API for questions with a fixed set of answers. OpenAI says it made its GPT-6 Luna-based Decisions API generally available on October 22. Developers define the allowed answers for a classification, choice, or score. OpenAI says the Decisions API can run up to ten times faster than GPT-6 Luna through the Responses API. That is a vendor claim for bounded tasks. It does not establish the same gain for open-ended reasoning or an entire agent workflow. The API now accepts up to 400 questions in one request. If one question depends on an earlier answer, though, it needs a separate request. Teams need to measure accuracy and speed on real decisions, including cases that need human review. A fast wrong answer can still send work to the wrong place. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.