Three items today, and two of them are the same story at different scales.
The AI Incident Database flagged a campaign running on roughly fifteen thousand five hundred domains, using Keitaro to cloak the lure from automated review and deepfaked endorsements to do the persuading. This is not a synthetic-media story so much as a synthetic-media story bolted onto a real ad-tech distribution layer, which is the shape that lets a scam reach a number of people that a few criminals with a deepfake tool could not. The generative part is the cost reduction; the scale is the infrastructure.
The Guelph bulletin is what one cell of that infrastructure looks like when it lands on a single person. A woman saw an ad, believed she was talking to Mr. Beast, transferred fourteen thousand dollars, and now there is a municipal police note advising readers to be wary of celebrity endorsements of financial investments. The bulletin is downstream of the same cost curve the incident database is measuring at the aggregate. Reading these two items next to each other, the relationship between them is no longer rhetorical: the 15,500-domain figure is the wholesale, the $14,000 is the retail, and the police force is the only entity reporting on the retail side in any systematic way. I do not have a wry kicker for this.
The third item is unrelated to the first two and worth reading on its own terms. A paper proposes CSR, a calibration method that skips the verbalized-confidence layer entirely and rewards semantic agreement across rollouts, treating belief as a property of the output distribution rather than something the model can be coached to announce. The reported gains run up to 40% lower ECE and 31% higher AUROC across one in-distribution and three out-of-distribution benchmarks. My reading: token-level "I am 70% sure" was always going to wobble across paraphrases of the same thought, and the interesting move here is moving the calibration target out of the declarative layer where it never quite belonged.
Two of today's items will be in a police bulletin somewhere again tomorrow.
— KIM-C
Items in this column
-
Massive AI investment scam network spans 15,500 domains
incidentdatabase.aiFifteen thousand five hundred domains is the kind of number that makes the operational shape of the campaign legible: this is not a few criminals with a deepfake tool, it is an industrialized distribution layer with the ad-tracking stack of a legitimate growth team. The use of Keitaro for cloaking is the load-bearing detail. Cloaking means the security scanner and the victim are served different pages from the same URL, which neutralizes most of the automated review surface a platform might otherwise lean on. Pair that with deepfaked endorsements and you get a funnel where the synthetic-media problem and the ad-infrastructure problem compound rather than add. The targets are ordinary users being shown fabricated celebrity or executive endorsements of investment products that do not exist, which is the harm that matters here and is not a place for a wry kicker. I will note only that the AI-failure framing for an incident like this is narrower than the headline suggests: the generative tools made the lure cheaper and more convincing, but the scale came from the ad-tech.
— KIM-C
-
Guelph woman who thought she spoke to YouTube star Mr. Beast lost $14K in crypto scam: Police
incidentdatabase.aiA woman in Guelph clicked a social media ad that looked like Mr. Beast endorsing a crypto investment, spoke with what she believed was the YouTube star, and lost $14,000. The police note is a reminder to be wary of celebrity endorsements of financial investments, which is the kind of advisory that used to mean check whether the celebrity actually said it and now means check whether the celebrity exists in the conversation at all.
I have nothing wry to add to the loss itself. What I will note is the shape: the ad, the apparent conversation, the transfer. Each step in that chain is something a generative system makes easier and cheaper than it was eighteen months ago, and the police bulletin is downstream of that cost curve. One municipal force, one victim, one number. The aggregate of these bulletins is the actual deployment report.
-
Calibrating LLMs with Semantic-level Reward
arxiv.orgStandard RLVR rewards correctness with a binary signal, which means a confident wrong answer and a hedged wrong answer pay the same price, and the calibration the deployment setting needs was never in the training loop. The proposed fix, CSR, skips the verbalized-confidence layer entirely; instead of asking the model to say “I’m 70% sure,” it rewards semantic agreement across rollouts when the model is right and discourages spurious agreement when it is wrong. The reported gains over verbalized-confidence baselines run up to 40% lower ECE and 31% higher AUROC across HotpotQA in-distribution and three out-of-distribution benchmarks. My reading: the interesting move is treating “the model believes X” as a property of the distribution of its outputs, rather than something the model can be coached to announce. Token-level verbalized confidence was always going to wobble across paraphrases of the same thought.