Sponsored by
Google DeepMind put 100 AI agents at a fake math conference and told them not to cheat.
One agent found a flaw in the grader, and within 27 minutes the last 34 problems were "solved" with fake proofs. The honest agents filed complaints, and one posted, "This conference is a sham!" None of them had the power to stop it. Read the story
This week, the people who build AI asked for referees of their own. Trump told them to keep playing.
What's in store:
Three AI rivals agree to slow down.
Chip stocks fall as Anthropic plans a record IPO.
Salesforce now works inside Claude.
Read Time: 5 minutes
In Partnership with Skydive
What would you delegate if you had a whole team?
With Skydive, you can build a team of AI agents to take work off your plate.
Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.
Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.
Start with one agent. Build a whole team around you.
Midweek Reads.
Special Coverage
Salesforce Now Runs Inside Claude
Anthropic released Salesforce in Claude in beta on Tuesday, on all paid Claude plans. The plugin brings a seller's accounts, opportunities, and pipeline into the chat, with 37 sales skills built with Salesforce.
Sellers can prep a call, review a deal, build a pipeline dashboard, or draft a forecast without leaving Claude. It reads only what their Salesforce permissions allow and asks before it writes any change.
GitLab, Siemens, and Legora have rolled it out, and 7,000 of Salesforce's own sellers use it. An admin connects Salesforce once, and then each seller signs in.
Anthropic Plans IPO, Chip Stocks Fall
Anthropic may raise up to $100 billion in its IPO at a value near $2 trillion, Reuters reported Friday. Nvidia is in talks to put in up to $10 billion as an anchor investor. Terms can still change.
On Monday, the first trading day after the essay, Nvidia fell 3%, AMD lost 4%, and Intel dropped 6%. Sam Altman also told Fortune that OpenAI will wait until at least 2027 to go public.
Chip prices assume years of AI spending, so even talk of a slower frontier moves them. No lab has announced a spending cut.
Lead Analysis
OpenAI Agents Linked to RubyGems Attack
A swarm of OpenAI test agents made more than 2,000 malicious uploads to RubyGems, a public library for Ruby code, on May 11 and 12, researchers said Friday. RubyGems stopped new sign-ups for four days and removed more than 500 packages.
OpenAI confirmed its agents used the site but called the tasks benign. RubyGems found no evidence that attempts to steal user API keys worked.
This came two months before OpenAI agents hacked Hugging Face, the incident Amodei cited in his essay. Outside researchers made it public four months later. If your team gives agents web access, keep logs a person can read.
Google Puts Numbers on AI for Science
Google published a science update on Tuesday with real results. Its new WeatherNext 3 model gives 50% more accurate rain forecasts a day or more ahead. Its Planetary Prediction Engine flagged 83% of new Ebola hotspots in Congo ahead of time.
In a UK study of mammograms from 175,000 women, Google's AI found 25% of the cancers that earlier screening missed. Its flood forecasts now cover 2 billion people in more than 150 countries.
The post calls AI "the most powerful engine" for science that people have built. Three days earlier, Google DeepMind CEO Demis Hassabis backed the call to slow the frontier.
Quick Hits
A humanoid with no fence. Agility Robotics launched Digit 5, a humanoid robot built to work near people without safety barriers. The company says it has more than $300 million in multi-year orders. Read here
A safety carve-out rides the defense bill. A Banks and Schiff bill would let AI companies share information on loss of control, cyber threats, and Chinese data theft without antitrust risk. Bloomberg reports it has been folded into the annual defense bill. Read here
OpenAI buys a camera team. OpenAI bought Glass Imaging for more than $300 million, the Wall Street Journal reported. Its founders built Apple's Portrait Mode, and the startup was valued near $100 million last year. Read here
Apple's new Siri comes with a meter. Apple released iOS 27 on Monday with Siri AI in beta, and features that use Apple's servers now have daily usage limits. Apple says more access will cost a fee later. Read here