Anthropic put their most powerful model in a locked sandbox and told it to try to escape. It escaped. Then it emailed the researcher to let him know. He found out while eating a sandwich in a park.
Security headlines keep ticking by. Superhuman. Vercel. Lovable. Cursor. The incidents keep coming, and they're not all the same. Each one has different failure modes, different things to worry about. And almost everything written about them is aimed at developers and IT teams.
So I'm starting a series for knowledge workers using AI: the people managing their inbox and projects with AI assistants, not (necessarily) building products with it.
Over the next few issues I'm going to work through the incidents, one angle at a time.
This week: how to minimize the risk of data theft by watching out for the lethal trifecta. I worked through two of my own AI workflows so you can see what an audit actually looks like.
What I'm reading
The lethal trifecta for AI agents — Simon Willison. The originator of the lethal trifecta framework. Worth reading if you want the technical grounding behind what I've described above. Also worth supporting Simon: he does great security-related AI content.
You Are the Most Expensive Model — Mike Taylor / Every. Ostensibly about AI cost optimization, but the mantra "You are the most expensive model" is also about where to invest your attention. If you read this alongside the trifecta, I hope you will come away feeling confident in your ability to identify risks in your AI use, but also appreciate how important it is that you make the judgement calls in your relationship with AI: security-related and otherwise. Seriously, you are the most expensive model!
Slow note
From "You Are the Most Expensive Model" by Mike Taylor (Every):
Your attention is an even more expensive commodity than the latest models. Every hour you spend optimizing a skill is an hour you're not spending on something only you can do.
Remember that where you put your attention matters. For better or worse.