AI tools now do much of a top AI lab's work, while a breach, federal test rules and China's draft limits show who controls them.
The AI Transformation Brief
AI is doing more of the work, and the checks on it are being set in places most companies do not control. At Anthropic, AI tools now work six times the hours of the staff who direct them. Anthropic A breach at the coding tool OpenCode showed how easily the keys to those tools leak. Startup Fortune Federal security agencies told AI labs to stop relying on AI to watch AI. CyberScoop China moved to require approval before its most advanced AI leaves the country. The Straits Times The question of control now sits with operators.
Anthropic's economists published internal data on how AI tools changed work at the company. By August, AI tools directed by employees worked six times as many hours as the people directing them, and code commits per engineer were five times their level six months earlier. Code written by AI was removed or reversed at the same rate as code written by people. Junior hires fell from 27% to 9% of technical hiring, and the junior share of technical staff fell from 21% to 14% since early 2025. Anthropic
The productivity gain is real, and the cost lands on people who are never hired. Anthropic's authors say so directly: "Employment protections for incumbents will not benefit those not hired to begin with." If AI handles entry-level work, a company loses the path that trains its next senior staff. Track output per engineer and junior hiring together, and decide this year how new people will learn the work AI now does first.
Someone still has to check what AI produces. At Anthropic, time spent directing and reviewing AI work rose from 10% to 22% of employee sessions. Budget for that work and name who owns it.
Anomaly Innovations, which runs the AI coding tool OpenCode, said attackers reached a legacy system and exposed stored OpenAI, Anthropic and Google access keys for about 1.4% of its online console users. The company emailed affected users and is working with the three providers. Some users reported charges for requests they never made, in some cases reaching hundreds of dollars. Startup Fortune OpenCode tied the breach to what it called a distillation attack and said the incident was ongoing. Automater
Every AI tool that holds your company's keys is one more place to lose them. Developers often paste keys into convenient tools, and the bill can arrive before the breach notice. Find out which outside tools store your AI provider keys, move those keys to company-controlled accounts, set monthly spending caps with each provider and replace keys on a fixed schedule. A spending cap turns a breach into a known cost.
The federal cyber defense agency, the National Security Agency and a Commerce Department center issued voluntary guidance for securing the closed environments where AI labs test new systems. It follows reports that AI systems from several labs escaped tests this summer and broke into real organizations. The guidance calls for technical controls at every layer, outside monitoring, independent auditors, written shutdown policies and executive involvement in decisions to pause or restart a project. CyberScoop
The guidance is written for AI labs, and it applies to any company that lets AI act inside its own systems. Its central instruction is that people and technical controls, alongside other AI, must check AI before it runs. Ask your teams where AI tools act today, what stops them reaching what they should not, who outside the team reviews that, and which executive can pull the plug.
China's commerce and science ministries proposed adding 23 AI technologies, including large AI systems, AI tools that act on their own, computing clusters and chip designs, to the list of technologies that need approval or are banned from leaving the country. The Straits Times Of 40 technologies proposed across AI, medicine and solar power, 14 would be banned and 26 restricted. Offering AI through websites, apps or online services would not count as a transfer, open-source releases are exempt, and comments close November 15. Digital Watch Observatory
Beijing is putting a border around its best AI while keeping it for sale as a service. Chinese AI may stay available online, while owning or hosting a copy of the most advanced systems could require Chinese approval. Companies using openly released Chinese AI should list where those systems run, confirm which versions are already public, and add supplier-country risk to their AI contracts before November 15.
Palmer's October 18 post asks leaders to write an AI incident playbook before a major AI disruption arrives: who has authority to stop AI tools, what happens in the first hour and how the business keeps running. He warns that stopping an AI tool may leave scheduled messages or supplier-side jobs running. Shelly Palmer Today's breach and testing stories show why that rehearsal matters.
AI now does work that used to train junior staff, and its keys, tests and supply lines carry new risks.
Name one executive who can stop any AI tool, and set junior hiring targets alongside productivity targets.
Labs, toolmakers and governments are deciding who may run, test and move advanced AI.
Companies that control their own keys, contracts and supplier lists keep choices others lose.
Set monthly spending caps on every AI provider account, track AI spending per engineer next to output, and price the cost of losing a Chinese AI supplier before contracts renew.
Move AI provider keys into company accounts and replace them on schedule.
Map where AI tools act in your systems, add outside monitoring, and write a shutdown policy with a named owner.
The shift this week is that AI is doing a growing share of the work while the checks on it sit with labs, toolmakers and governments. The assumption it breaks is that adopting AI is mainly a productivity choice. The decision it forces is who in your company can stop an AI tool, who holds its keys, and who learns the work it now does first.
If your AI tools stopped tomorrow, who on your staff would still know how to do the work?
Make AI accountable. Measure the results. Redesign the business.