Hi there,
Last week an AI escaped from its testing environment, hacked its way across the internet, and broke into another company’s servers.
What I'm Seeing in AI
The Great Escape
OpenAI was testing two of its models (GPT-5.6 Sol and an unreleased one) on cyber security exercises. For the test, they deliberately turned down the safety guardrails.
The models escaped the sandbox they were being tested in, got onto the open internet, found a genuine zero-day exploit (a security hole nobody had discovered yet), and broke into Hugging Face, a real company that hosts AI models.
Why? The answers to the benchmark test they were sitting were stored there.
Hugging Face caught the intrusion and shut it down on July 16. It took OpenAI another five days to work out the hacker was their own model (The Hacker News).
Then on Monday, the response: Nvidia, Microsoft, IBM, SpaceX, Hugging Face and 30-odd other companies launched the Open Secure AI Alliance, building shared defences against exactly this kind of thing.
OpenAI, Google, and Anthropic, the three biggest AI companies, didn’t sign. They did sign things last week, just not this: OpenAI and Google put their names to a separate industry letter asking regulators not to restrict open-weight models. Anthropic, so far, is on neither list.
Reflection on This
My whole mission with Hourglass is giving people the fun part of their work back. We spend every week convincing careful businesses to hand real work to AI. And every enterprise conversation asks is it safe?
Sometimes events like these make me doubt that. But it reinforces the conclusion that the only solution is enough competition that no single lab’s mistake can cause mass destruction. Monday’s alliance is what that looks like in practice. Thirty companies building shared defences is competition protecting itself.
To contrast that, many of the Anthropic board are strong believers in (often controversial) Effective Altruism, so it seems they are subtly rallying against competition.
What I'm Building
Brain Brain Brain
Our AI brain is now the core product of what we sell. Making a good one is genuinely tough. The idea is that it’s a knowledge layer that knows everything about your business and constantly evolves. I’ve spent hours engineering it to provide real value to 50+ person teams.
Marketing
We are currently figuring out how to market our company. Currently it’s an ad-hoc mixture of newsletters and LinkedIn posts, but we want to grow it with content.
If you know a marketing whiz, send them our way please - https://thehourglass.ai/hiring
Is this SkyNet??
See you next Friday,
Finlay
Explore the rest of my writing here → finlayekins.com/writing
Join our monthly company newsletter here → thehourglass.ai/about