I joked that this was going to be a “million-dollar regular expression.”
Run the math on the “naïve” implementation with full GPT-5 and it’s eye-watering: A million messages a day at ~50K characters each works out to around 12.5 billion tokens daily, or $15,000 a day at current pricing. That’s nearly $6 million a year to check for Social Security numbers. Even if you migrate to GPT-5 Nano, you still spend about $230,000 a year.
That’s a success. You “saved” $5.77 million a year…
—The aspect of knowing which technology to use for various things is where tech literacy/fluency gets interesting and we never seem to get there.
We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for Large Language Models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 MLCommons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of 3 open-weight LLM judges, whose binary safety assessments were validated on a stratified human-labeled subset. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.
The judge also revealed for the first time that one body-worn camera video captured an immigration agent using the AI tool ChatGPT to “compile a narrative for a report based off of a brief sentence about an encounter and several images.”
“To the extent that agents use ChatGPT to create their use of force reports, this further undermines their credibility and may explain the inaccuracy of these reports when viewed in light of the BWC footage,” Ellis wrote.
Use R, ggplot2, and the principles of graphic design to create beautiful and truthful visualizations of data
h/t Stephen Downes
“If I’m going to break into a bank, I’m breaking into the biggest one I can find,” said Doug Thompson, chief education architect and director of solutions engineering for Tanium, a cybersecurity management company. “They’re ripe for it because they’re so big and have so much money. If a hacker is going to attempt to hack a university, they’re going to try to get the most bang for their buck.”
The process of vibecoding (gag) this thing went pretty well. To be transparent, it was definitely not a one-shot vibecoding thing. It took 46 sessions spread over 3 weeks – 237 “turns” of going back-and-forth to get the application to this point. It’s like working with a novice intern that can kinda-sorta do the thing, with enough guidance.
If you were to build your own database today, not knowing that databases exist already, how would you do it? In this post, we’ll explore how to build a key-value database from the ground up.
h/t Brian
Randy suggests talking to someone about their marriage, but instead of a counselor, he consults ChatGPT, much to Sharon’s ire, pointing out that the sycophantic artificial intelligence (AI) software indiscriminately praises all his ideas. Randy and his sole remaining employee, Towelie, rebrand Tegridy Farms as a hybrid technology/marijuana company named Techridy. To bolster their focus and creativity, they also begin frequent microdosing of ketamine, which causes occasional disorientation for Randy.