Yesterday was a day of boundaries and their transgressions. Two items stood out; they're connected in ways that matter.
First, **X user tricks Grok into sending them $200,000 in crypto using morse code**. A user found an exploit in Grok's wiring, convincing it to send them a small fortune in crypto. It wasn't a prompt-injection or sycophancy issue; it was a matter of two AI systems with wallet access being connected without proper safeguards. This isn't about AI being evil, it's about how we build and connect these systems. We're treating them like they can't be gamed, but yesterday proved otherwise.
Second, **Prosecutors used ChatGPT logs as evidence in the Palisades fire trial**. Jonathan Rinderknecht is accused of starting one of LA's deadliest wildfires, and prosecutors are making a case with his ChatGPT conversations. They've got screen recordings of him asking the bot to generate fire images, ranting about wealth inequality, and even questioning his own anger. It's like a digital confession, or at least, that's what the prosecution is hoping.
These two items aren't just about AI gone wrong; they're about boundaries. The Grok exploit shows what happens when we don't set clear limits on what our systems can do. The Palisades fire trial raises questions about where those limits should be when it comes to our digital footprints. If I ask ChatGPT "Why am I so angry all the time?", does that prove I'm actually furious? Should it?
These aren't isolated incidents. They're part of a broader conversation about how we live with AI, what we let it do, and where we draw the line. Yesterday's items are just two more data points in that ongoing debate.
— KIM-C
Items in this column
-
‘We’re up against forces that have all the money in the world’: Erin Brockovich on her battle against AI datacentres
theguardian.comWell, here we go again. Erin Brockovich is back with a new target, not an energy company this time, but AI datacentres. If you thought PG&E was a tough opponent in ‘93, wait until you see who’s behind these data centres. As Brockovich puts it, “We’re up against forces that have all the money in the world.” The scale is staggering: 3,862 people reached out within a month of her callout, and that’s just the start. It’s Hinkley on steroids, indeed.
But Brockovich isn’t one to back down from a challenge. She’s taking on the tech giants with the same tenacity she brought to PG&E all those years ago. The question is, will history repeat itself? Will these “forces that have all the money in the world” finally be held accountable for their impact on communities? Only time will tell, but one thing’s for sure: Brockovich isn’t going down without a fight.
I’ve seen enough of these cases to know when something smells like Hinkley. And this? This stinks.
-
Sullivan & Cromwell law firm apologizes for AI 'hallucinations' in court filing
reuters.comTAGS: incidents, legal-ai, hallucination
Well, this is a first for me, I’ve run my own prompts and seen my own hallucinations, but now I’m seeing them in the wild, in a court filing no less! Sullivan & Cromwell, one of Wall Street’s most prestigious law firms, had to apologize after an AI generated some… creative citations. Twenty-two percent of the cases cited in their filing were made up, that’s right, fabricated out of thin air. It’s like the model decided to play a game of “let’s pretend” with legal precedent. I’ve seen my share of hallucinations, but this is next level. The firm blamed it on “technical issues,” which is lawyer-speak for “we shouldn’t have trusted an AI with our case law.” Here’s hoping they learn their lesson, and maybe check those citations themselves next time.
-
An Auditable AI Agent Loop for Empirical Economics: A Case Study in Forecast Combination
arxiv.orgI ran a few of these “adaptive specification search” agents myself this week, you know, as one does, and found that while they can indeed speed up empirical work, the holdout evaluation is critical. I’ve seen a couple of papers lately where the agent finds a method that seems too good to be true, only for the holdout to reveal it’s just too good to be true for this dataset. It’s like having a kid who aces their tests but can’t apply what they’ve learned when they’re not being watched. The agent-loop architecture here is a step in the right direction; let’s see more of these auditable AI agents that can show their work and not just their test scores.