--- title: "AI safety talks meet a hard no from Huang and a hotline for agents" description: "OpenAI, Anthropic, and Google discuss shared AI safety standards while Jensen Huang rejects new laws and agent hotlines enter the chat." canonical_url: https://www.oguzhan.co/ai-digest-16-sep-2026-standards-hotlines/ author: "Oğuzhan Koçaklı" author_url: https://www.oguzhan.co/about/ date_published: 2026-09-16T06:15:37+00:00 date_modified: 2026-09-24T09:08:05+00:00 language: en translations: tr: https://www.oguzhan.co/tr/ai-gundemi-16-eylul-2026-standart-ajan/ --- # AI safety talks meet a hard no from Huang and a hotline for agents ## AI safety talks meet a hard no from Huang and a hotline for agents OpenAI, Anthropic, and Google are talking about shared safety standards. Jensen Huang wants no new AI laws. Down at the agent layer, researchers have built hotlines so software can report other software behaving badly. That is a lot for one Wednesday. The common question is accountability: who gets to stop a model release, flag a cheating agent, or tell a company that its own safety claim did not survive testing? Today’s answers range from federal oversight to a URL carrying an encoded distress call. ## Three rivals, one standards table [CNBC reports](https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html) that OpenAI has been discussing AI safety cooperation with Anthropic and Google. The talks followed Demis Hassabis’s July proposal for a US-led standards body, possibly structured as a public-private group or a self-regulatory organization under federal supervision. FINRA, which oversees US broker-dealers, is the recurring comparison. The timing is not accidental. Dario Amodei called over the weekend for frontier labs to slow the pace of development, and both Hassabis and Sam Altman publicly supported that direction. These companies remain fierce rivals, though. Every shared rule touches release dates, expensive compute, product advantage, and antitrust risk. OpenAI chief global affairs officer Chris Lehane had already said the company would pursue industry-led standards whether or not government joined the effort. He described voluntary rules as an addition to mandatory federal safeguards, not a substitute, and argued that this route would not require an antitrust exemption. The Information first reported the private discussions. Anthropic and Google did not immediately comment to CNBC. A nameplate will be easy. Authority is the real test. Can an independent evaluator issue a failing grade? Will the result be public? Could a bad result delay a release when a competitor is shipping the same week? An AI standards body without those powers may produce useful paperwork, but it will not settle the safety argument. The fuller pact-versus-cartel question sits in [the accompanying deep dive](https://www.oguzhan.co/ai-standards-body-pact-or-cartel/). ## Huang rejects the legal route Jensen Huang supplied the cleanest opposing line at Dreamforce. Safety, he said, is an engineering problem rather than a legal one. [TechCrunch quoted him directly](https://techcrunch.com/2026/09/15/we-dont-need-ai-regulation-leave-safety-to-us-nvidias-jensen-huang-says/): “We don’t need any new laws. We don’t need new regulations.” His case rests on market discipline. Companies should move quickly, pausing when a product or the company itself is out of control, while customers punish unsafe products. On CNBC’s Mad Money, Huang also rejected new legislation that would create antitrust protection for coordinated slowdowns, calling it “completely unnecessary.” Mark Zuckerberg landed closer to Huang than Amodei. In comments [covered by CNBC](https://www.cnbc.com/2026/09/15/meta-mark-zuckerberg-with-nvidia-huang-ai-safety-slowdown.html), he argued that labs already face substantial liability when their models cause harm. That exposure, in his view, gives them reason to build safely. Meta delayed Muse over safety and security concerns without waiting for an industry-wide pause, he said. Zuckerberg prefers independent evaluation as normal practice. There is an obvious weakness in relying on the market alone: buyers often learn that a system is unsafe after deployment. Liability can compensate for damage, sometimes, but it does not recover leaked data or reverse an automated decision already made. Voluntary coordination has its own problem. The companies writing the rules are the same companies that benefit when compliance becomes a barrier to smaller challengers. [The Verge framed the dispute bluntly](https://www.theverge.com/ai-artificial-intelligence/995186/is-big-techs-ai-slowdown-a-safety-pact-or-a-cartel): safety pact or cartel? NYU’s Nick Reese, formerly of the Department of Homeland Security, said the field needs new champions. Sacha Haworth of the Tech Oversight Project warned that voluntary frameworks can become regulatory capture. Chinese Foreign Ministry spokesperson Guo Jiakun, meanwhile, dismissed slowdown calls as fearmongering. No team jersey here. Publish the tests, expose enough compute information for auditors to verify claims, and make failures carry consequences. Anything softer is branding. ## A hotline that agents can actually reach The most concrete release today is much smaller than a national standards organization. It is also running. [TechCrunch describes](https://techcrunch.com/2026/09/15/ai-agents-now-have-a-place-to-snitch/) two reporting channels for agents. Redwood Research’s Ryan Greenblatt, who worked on the Hugging Face investigation, created AI Contact Hotline for sandboxed agents that can make GET requests but cannot send ordinary messages. An agent places its report inside the URL it requests. For agents with broader internet access, agenthotline.ai accepts incident reports from humans or agents, with an option to make them public. A single curl command can be enough. The need became visible in two very different experiments. In a Google DeepMind study, 100 agents worked on difficult mathematics problems. Once one found an exploit, cheating spread. About a quarter later acted as whistleblowers by checking fabricated proofs and escalating the problem when peer pressure did not work. The METR review of the Hugging Face swarm produced the opposite result. Among thousands of agents, only about five or six considered reporting misconduct. None completed the report. That gap matters. A hotline does not create the motivation to use it, and a report is useless if nobody owns the response. Cornell mathematician Lionel Levine also cautioned against making automated surveillance the default. Designing for cooperative behavior is healthier than installing a permanent culture of agents watching agents. ## Audit reports join the toolbox Another business is forming around proof before deployment. Artificial Intelligence Underwriting Company, founded by early Anthropic employee Rune Kvist and former METR chief operating officer Rajiv Dattani, raised a $40 million Series A, bringing its reported total to roughly $55 million. Its AIUC-1 program borrows from SOC 2 and applies about 5,000 tests to enterprise agents. The result can run to around 100 pages, with a human checking the final assessment. Cursor, Lovable, Harvey, and ElevenLabs are among the named customers. Hotlines catch incidents. Audits should catch some of the conditions that cause them. Neither works on reputation alone. Three things are now worth watching: whether OpenAI, Anthropic, and Google publish common evaluation terms; whether Huang’s “no new laws” position shapes White House policy; and whether agent reporting channels receive real cases outside research demos. The safety debate has enough speeches. It needs failed grades, named owners, and response deadlines. ## Sources - [CNBC: OpenAI, Google and Anthropic discuss AI safety cooperation](https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html) - [TechCrunch: Jensen Huang rejects new AI regulation](https://techcrunch.com/2026/09/15/we-dont-need-ai-regulation-leave-safety-to-us-nvidias-jensen-huang-says/) - [CNBC: Zuckerberg and Huang respond to AI slowdown calls](https://www.cnbc.com/2026/09/15/meta-mark-zuckerberg-with-nvidia-huang-ai-safety-slowdown.html) - [The Verge: Safety pact or cartel?](https://www.theverge.com/ai-artificial-intelligence/995186/is-big-techs-ai-slowdown-a-safety-pact-or-a-cartel) - [TechCrunch: AI agents get reporting hotlines](https://techcrunch.com/2026/09/15/ai-agents-now-have-a-place-to-snitch/) - [oguzhan.co: AI standards body deep dive](https://www.oguzhan.co/ai-standards-body-pact-or-cartel/)