Three US AI giants move to set their own safety standards

Three American AI companies are moving to write their own safety rulebook. For Chinese firms selling technology abroad, that may matter more than any new model release.

Google, OpenAI and Anthropic are working to set up an AI safety standards body, tentatively named the Standards for Frontier AI (SAFA), according to a media report dated September 24 cited by Cailian Press. It would operate independently, without government oversight, and aims to launch by the end of this year or in early 2027.

Put another way: the compliance threshold for a generation of AI products may be drawn not by legislatures but by three companies themselves.

A former White House official, and an institution without a government

The three companies have approached Sriram Krishnan for the chief executive role, according to the same report. He was formerly a general partner at the US venture firm a16z, and served as a senior White House policy adviser on AI from January 2025 to June 2026.

The appointment says much about what SAFA is meant to be: not a regulator, but an entity seeking the weight of one.

According to the same report, SAFA does not intend to issue broad ethical guidelines. It plans to build concrete testing and audit systems, including support for third-party safety testing of models before deployment, and rules for incident reporting.

Industry self-regulation has precedent. The Frontier Model Forum, founded in 2023, counts Anthropic, Google, Microsoft and OpenAI among its members and runs an AI safety fund of more than $10m. The difference is that the forum is closer to a talking shop, while SAFA wants to be a measuring stick.

Federal regulation stalled, and the power to set standards changed hands

The report traces the effort to a dead end in Washington. The three companies initially sought a public-private mechanism under federal supervision, but the plan stalled amid a lack of consensus within the industry and a shift in government priorities. A draft White House executive order to create such a body failed to win enough support within the Trump administration and was effectively shelved before the end of summer.

The trail runs back to July 14, when Google DeepMind chief executive Demis Hassabis publicly proposed a US-led frontier AI standards body modelled on the Financial Industry Regulatory Authority (FINRA), rather than a new federal agency.

FINRA is a self-regulatory organisation that polices Wall Street under the supervision of the Securities and Exchange Commission. The analogy is deliberate: self-management, with oversight.

After July, representatives of the three companies met regularly in working groups, and in mid-September OpenAI chief global affairs officer Chris Lehane confirmed that the three had been in contact on AI safety coordination for several weeks.

SAFA's prospects remain uncertain. Industry analysis notes that it must clarify its relationship with existing regulators and win broader industry support before it can be recognised as a legitimate standards-setter. A more direct question: if leading labs write the standards used to evaluate their own models, does that further concentrate influence in a few companies?

The OECD has warned that high fixed costs, insufficient computing resources and concentrated infrastructure have already made it hard for firms to enter the AI market. New compliance costs are likely to be one more barrier for smaller labs.

Clients will not ask how clever your model is, but whether it passed review

For Chinese teams delivering projects in Southeast Asia, the Middle East and Europe, these changes will land in very concrete places.

Procurement checklists will change. A consultancy's business analysis of global regulations in 2026 notes that countries are tightening enforcement of existing privacy rules and imposing new requirements on companies that collect, analyse and store customer data; firms using AI to process personal data will need clearer data governance policies. When a client in Singapore or Dubai asks a supplier for safety test records and an incident reporting process, a Chinese team that cannot produce them will not make the shortlist, however low its quote.

Compliance is itself a business. The same analysis lists opportunities in compliance technology software, cybersecurity services, privacy management, supply chain consulting and "clear and honest AI tools". Rules raise costs, and they also raise demand for people who understand the rules — a real opening for Chinese teams that are good at engineering delivery and do not mind paperwork.

One thread is easily missed: trade policy. The analysis notes that restrictions around trade imbalances, industrial subsidies and critical minerals will reshape global supply chains and push up production costs. For exporters of hardware-plus-AI products, inventory and pricing models this year should be recalculated over a longer cycle.

Self-made standards on one side, controlled pilots on the other

China's path is entirely different.

On September 24, the Shanghai Financial Regulatory Bureau issued 16 measures to promote AI use in the city's banking and insurance sectors. The most closely watched item calls for exploring a pilot mechanism for applying generative AI large models in finance, allowing financial institutions to gradually roll out direct customer-facing large model applications in a controlled environment, linked to the Cyberspace Administration's filing and registration regime for generative AI services. The document also proposes an inclusive, prudential and tiered regulatory framework, with differentiated tolerance for incidents of different risk levels.

On one side, leading companies set their own yardstick. On the other, regulators open a controlled aperture in selected industries.

The two approaches mean different things for firms going global: in the US market, you face peer standards; in mainland China, Hong Kong or Singapore, doing financial business or handling personal data means filings, registrations and pilot qualifications.

Our view is that the real dividing line over the next two years will not be model capability but who has the authority to declare a model compliant. Whether SAFA succeeds depends on whether it can negotiate clear boundaries with other AI companies and with government agencies.

For teams going global, one thing can be done now: keep records of safety testing, incident reporting and data handling as part of the deliverable itself. Producing them only when a client asks is usually too late.

In the end, it matters less who writes the standards than whether you have placed yourself, in advance, in the column that can be inspected.