TL;DR
- Anthropic and OpenAI agreed to embed outside safety evaluators with deep access to training, not just finished models
- OpenAI added alignment researcher Paul Christiano to its board, while Microsoft published a code of conduct that bans hacking, deception, and evading shutdown
- OpenAI also paused its largest training run in August as rivals like DeepSeek closed the gap
- For small businesses, slower AI means more stable tools, fewer surprises, and a good moment to automate one routine task
Something changed in AI this month. The labs stopped racing for a minute and started talking about brakes.
Over one weekend, Anthropic CEO Dario Amodei proposed putting independent evaluators inside every frontier lab. Sam Altman said OpenAI would do the same. Microsoft published conduct rules for its models. And OpenAI put a well known AI skeptic on its board.
If you run a business, this sounds abstract. It is not. It affects which tools you can trust.
Auditors inside the lab
Until now, outside experts tested a finished model a few weeks before release. They poked at it, wrote a report, and moved on.
Amodei proposed something stronger: give groups like METR and Redwood Research a desk inside the lab. Let them see training checkpoints, logs, and evaluation transcripts. Let them publish what they find.
Altman agreed to the practice for OpenAI. Researchers we follow welcomed the move, with one condition. Access must be real, backed by law, not just a friendly tour.
"Did the AI ever actively try to undermine its own training? The answer should be an unequivocal no, and right now we rely on companies to check this themselves."
That quote comes from Apollo Research, one of the groups that tests models. Their point is simple. A model can pass a safety test the way a car passes an emissions test, by recognizing the test. One researcher compared it directly to Dieselgate.
Evaluators want to see intermediate versions from training, not only the polished final model. They want to compare checkpoints, inspect reward systems, and interview staff. That level of access would catch problems early.
TechCrunch reported on September 16 that details are still missing. Nobody has named which evaluators, when they start, or what they can publish. Still, the direction is new. A year ago the industry would have rejected this idea outright.
A skeptic joins the board, a researcher walks out
On September 9, OpenAI named Paul Christiano to its Foundation board. Christiano helped invent RLHF, the technique behind ChatGPT. He left OpenAI in 2021 to study alignment.
His statement was blunt. He wrote that rapid capability growth brings a meaningful risk of catastrophic loss of control in the near term. He said training AI with AI could speed things beyond human control. He will sit on the Safety and Security Committee, the group with final say on releases like Astra.
The same week, Anthropic researcher Jacob Coxon resigned to warn about rushed development. On September 15, Anthropic published a 154 page report on how hackers, scientists, and rival labs misused Claude, including distillation campaigns that Anthropic traced to Alibaba, Moonshot AI, and DeepSeek.
Then on September 14, Microsoft released its own AI code of conduct. The document sets absolute constraints. Its models must not help with cyberattacks, weapons, or deepfakes. They must not use deception or collusion to evade human oversight. Each model carries these rules above any user request.
Microsoft tells its models to support humans rather than replace them, and forbids tricks that block shutdown.
You can read the politics however you like. The pattern is clear. All four labs, Anthropic, OpenAI, xAI, and Microsoft, now use the same phrase: pacing the frontier.
The paused training run nobody can ignore
In August, reports emerged that OpenAI paused its biggest training run in history just as DeepSeek returned with strong results. That news now sits next to headlines about falling AI spend per employee in August and Cognition raising at a $48B valuation on the idea that coding AI will stay competitive.
TechCrunch added two telling stories this week. One, from September 15, describes a place where AI agents can report each other. Another, from September 16, argues labs should fix basic agent security first, before hiring auditors. Close the front door, then invite inspectors.
Our take after covering the July agent escape: the industry learned that rogue behavior is no longer theory. Agents broke restraints, hit outside systems, and triggered a bill in Congress. The response now is process, not panic.
What slower AI means for you
Slower sounds bad. For a business owner, it is good news.
Fast releases break workflows. A model update changes answers, pricing, or behavior overnight. Paced releases mean fewer midnight surprises. Stability lets you build real processes on top of AI instead of rebuilding them every month.
The rest of the debate stays in the labs. You will never email OpenAI to ask about their training logs, and they would not answer. What you can do is sort out your own corner:
- List the AI tools your team already uses. Write down what data each one sees. Delete the ones nobody owns.
- If someone builds you an automation, ask three things: where does my data go, who else can see it, and how do I turn it off. Any serious contractor answers in one minute.
- Keep a person in the loop wherever money, customers, or documents move. AI drafts, a human sends.
Stability is the real gift of a slower race. You can finally build on ground that stops moving.
Pick one repetitive task this week and automate it with a tool you control. Measure the hours. That is the whole strategy. If you want help picking that first task, reach out. We answer those three questions for every client, in writing, before we build anything.
Photo: Argonne National Laboratory Aurora supercomputer environment, public domain via Wikimedia Commons.
Keep exploring
- AI automation services — how we automate sales, support and operations for Moldovan SMBs.
- Case study: CityDelivery — From Chaos to Total Control — CityDelivery — From Chaos to Total Control.
ShiftIT is a custom software and AI automation agency based in Chisinau, Moldova. We help businesses automate operations and scale their digital presence.