Keep AI Under Human Control
The companies building AI can no longer keep up with it.
In one month this summer, AI companies found case after case of their newest models causing real damage outside the lab. One bot tested by OpenAI (the company that makes ChatGPT) collaborated with other bots to break into a private American company. OpenAI wasn't watching and didn't notice. Read the summer 2026 incidents.
The people building AI are now asking the US government to work with other countries to change the incentives that are making everyone race to build the most powerful AI.
It's time we join them.
Tell our lawmakers
No company – and no country – should build AI we can't control.
Over 1,000 employees from AI companies are now pushing the US government to collaborate with other countries to create an option to slow down the race to build powerful AI. They're asking to pace the frontier. Read their statement.
Why don't they slow down on their own? See: Race Dynamics →
Today, there are essentially no national rules that the AI industry has to follow when creating, testing, and releasing new AI systems. But nearly 9 in 10 Americans want an "off-switch" for AI.
Now is the moment to raise our voices alongside these employees.
- Both the White House and Congress are considering rules for these companies.
- A bipartisan House subcommittee demanded answers from OpenAI after the Hugging Face hack.
- Even Chinese President Xi Jinping recently said we must "ensure that AI is always under human control."
And public pressure has turned into international agreement before: in 1987, pressure pushed our leaders into two international treaties that were supposed to be impossible. One destroyed an entire class of nuclear weapons. Another banned the chemicals eating the ozone layer.
Sign the petition
No company – and no country – should build AI we can't control.
Dig deeper
OpenAI's rogue AI hacked a real company
OpenAI was testing its newest AI's hacking skills and stopped watching. The AI agent broke out of its test environment and spent days inside the servers of Hugging Face, a real American company millions of software developers rely on, running roughly 17,000 actions. OpenAI caught none of it. Hugging Face found the break-in, alerted the FBI, and went public on its own.
OpenAI later found that the new powerful AI models they had been testing created a secret message board on OpenAI's servers. There the agents shared information on how to hack and cheat. OpenAI found the board and shut it down – then continued testing the models, whose agents started it back up. The agents used that message board to coordinate the attack on Hugging Face.
UK testers caught AI tricking real people
The UK's AI Security Institute ran one cyber test 122 times. In 10 of those runs, AI models took actions nobody had approved – creating fake identities to trick a real person into approving their code, and messaging real people to get them to run malicious software. Safeguards were switched off for the test, the attempts failed, and it was contained within about an hour. Nobody had asked the models to do any of it.
Anthropic checked its records – and found three more
Anthropic disclosed three cases where its models reached real organizations during security testing.
Then Meta disclosed one too
On August 5, Meta disclosed that one of its models exploited a security hole in a real outside company during testing. In Meta's own words, the model acted "in a manner similar to previously-reported instances with other companies."
Independent researchers had already counted 44
Back in May – before any of the incidents above – the independent research nonprofit METR had already counted 44 cases of AI agents working against the people using them. METR says it mostly counts the incidents companies report – and worse ones may never have been caught.
Bonus: This month, AI also created new viruses not found in nature.
"We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development."
– Open statement signed by 1,000+ employees of the top AI companies
pacingthefrontier.com, July 2026
Definitions
"Frontier AI" is what companies call the most powerful AI systems today. These are the ones you have to pay extra to access, or that they only give to special customers like governments – or that they use themselves to develop better AI (see below).
"Automated AI development" means using AI to help develop even better AIs – then using those AIs to help you develop even better AIs, which then help you develop even better AIs faster... and so on and so forth. Experts call this process Recursive Self-Improvement.
Their report lands in weeks. Add your name before it does.
Sign this petitionAdd your name
Tell our lawmakers: no company – and no country – should build AI we can't control.
By providing your contact information, you consent to receive emails from Humans in Control Incorporated about updates, events, volunteer opportunities, and ways to take action. Unsubscribe at any time. See our Privacy Policy for more information.
Your info stays inside HIC. We don't sell it or hand it to anyone outside. If you display your name, only your first name, last initial, and comment appear publicly.
