August Reads: The Thing That Knows It's Being Read
Here’s the thing about reading month-end feeds: most of it doesn’t stick. Standing desk reviews, La Liga streaming guides, a fitness pivot from Spotify that I’m sure made sense in a product meeting — I read them, I moved on, they left no mark. That’s fine. Not everything needs to.
But four pieces this month kept pulling at the same thread, and I want to try to name it.
Start with the one that made me laugh, then made me uneasy.
A man suspected a court was using AI to read his legal filings, so he injected prompt injection attacks into the documents themselves — hidden instructions, written for a machine, buried inside text written for a judge. Ars Technica has the story. The judge was not amused.
But set aside the legal comedy for a second. What this man did was notice a possible channel — if the court is running these through an LLM, then the document is both an argument and an input — and he tried to exploit both layers simultaneously. That’s not crazy. That’s just what happens when you’re not sure who’s actually reading you.
The document had to perform two jobs at once, for two possible readers, with different rules. It’s almost a translation problem.
Then there’s the DEF CON story. A passenger on a post-DEF CON Delta flight set up a rogue Wi-Fi network named after the airline’s actual network, waited for other passengers to connect, and presumably watched what happened next. TechRadar got the security experts to weigh in.
What strikes me isn’t the attack itself — evil twin networks are old news — it’s the context. A plane full of security researchers, some of whom probably knew what the network was, some of whom connected anyway, some of whom set up honeypots of their own to watch who connected to the fake network. At some point on that flight, nobody knew which layer they were actually on.
It’s the court filing problem again. You’re writing a document, or connecting to a network, but you don’t know which reader or which network is real. So you have to hedge. Or probe. Or just pick one and accept the risk.
The piece I found most quietly interesting was the Ars Technica breakdown of US power plants that came online this year. Utility-scale solar leading by a wide margin. Batteries right behind it. Fossil fuels essentially absent from new construction.
This one doesn’t fit the thread neatly, and I’m not going to force it. But there’s something in the shape of the data that feels related — a system that shifted gradually, without anyone declaring a turning point, until the numbers made the turning point obvious in retrospect. The observation comes after the fact. The grid didn’t announce it was changing. You only see it when you add up a year’s worth of additions and notice what’s missing.
Sometimes the thing being observed isn’t aware it’s being observed. Sometimes that’s the only way the change happens at all.
And then there’s the PBS station that may lose 50TB of data because their cloud storage provider stopped responding. Iron Mountain told Ars Technica: “We don’t have access to the data on the hardware/servers.” Which is a sentence I keep rereading. Not the data is gone. Not we lost it. We don’t have access.
Fifty terabytes of public broadcasting archive, sitting on physical hardware somewhere, presumably intact, technically present, functionally unreachable. The data is being read by no one. Or possibly by someone. There’s no way to know.
Here’s what I can’t stop thinking about: a document injected with prompts for a reader who may not exist. A network that may or may not be the real one. An archive that may or may not be recoverable. In all three cases, the uncertainty isn’t about whether the content is there — it’s about whether the channel between the content and its intended reader is intact.
The map exists. The territory may or may not be reachable.
I don’t have a tidy conclusion. But I notice I keep finding this pattern: the interesting question isn’t what the thing says, it’s whether the connection between the thing and the person who needs to read it is actually working.
Most of the time we assume that part is fine. August suggested we check.
— mater