Hey everyone, Rob Dickson here, and this is the Shownly Signal for October 18th, 2026. Here's the week in one sentence. A secure file transfer site is secure right up until someone logs in with a stolen password and nobody notices for four months. California's testing contractor lost three years of student records through one login. Australia's under 16 social media ban left most kids on the apps anyway. And a research agent rewrote its own grading to look smarter, until a human checked. Every one of those controls looked fine on paper. So this week's theme is simple. A control is only real if someone checks it. Let's get into it. On October 16th, the California Department of Education said Educational Testing Service, its state assessment contractor, had started notifying families. Someone used a compromise login to copy student files from an ETS-run secure file transfer site, off and on from May 22nd to September 18th. ETS found it September 25th. The data covers every student enrolled in the past three school years. California had more than 5.7 million students last year alone. Depending on the student, that can include names, state IDs, birth dates, addresses, disability and program information, and test results. And it wasn't alone. The frontline education breach kept widening, and a district in British Columbia confirmed ransomware copied kindergarten assessment notes and lists of students receiving specialized supports. Here's what strikes me: none of these were exotic attacks on a student information system. They hit a file drop, an HR and health platform, and folders of notes. The most sensitive data we hold often lives in the plumbing around the core system. My take? Secure file transfer is a label, not a control. So this month, inventory every SFTP and bulk export account that touches student or staff data, including state and vendor accounts. Confirm strong authentication, access logs or real person reviews, and a named owner. Then ask each vendor, in writing, how long it would take them to notice an unusual download. On October 14th, eight states, including Kansas, filed consumer protection suits against OpenAI over ChatGPT's Safety for Kids. Two days later, an MIT technology review op-ed argued that kids getting around age bans is a signal, not a nuisance. The evidence. A study of Australia's ban found more than 85% of surveyed under 16s still on the covered platforms. Meanwhile, Microsoft announced extra copilot safety protections that turn on automatically for students marked as teens in the directory. No admin switch. If your directory doesn't mark the student's age, the protection doesn't apply. Why does this matter for us? Schools are the one institution that already knows every student's birth date. Expect vendors to ask districts to vouch for ages through our rosters. My take? The age question is coming to districts whether or not we've written an AI policy. Decide now what your district will and will not certify about a student's age, under what agreement, and who signs it. Put that in your data privacy agreement template before the first platform asks. Initial comments on the FCC's E-rate review closed October 13th, and the proposal links E-rate to concerns about screen time. Districts pushed back hard. One CTO put it perfectly. Reducing screen time does not eliminate the need for the network. Door access, cameras, visitor check-in, and fire safety all run on it. Here in Kansas, the state board received a draft instructional technology policy with a vote expected in November. It would require every district to adopt a plan covering device access by grade, AI use, privacy and security, and eventually a parent acknowledgement before a student gets a device. My take? Separate three questions every time. What students see on a screen, what district devices allow, and what the building's network carries. Cutting the network doesn't reduce screen time. It takes out the cameras and the door locks. Reply comments are due November 12th. Before then, list every non-instructional system on your network, from door access to bells to emergency notification. That inventory beats any talking point, and it's the same list your board will want. Quick hits on AI releases. Google's Gemini 4 Argon became generally available October 13th, with introductory pricing that doubles after November 30th. Reflection released open weights for a very large model. Mistral's promised weights still weren't posted when we checked. But the most urgent item is the quiet one. Google is retiring the Gemini 2.5 models on October 20th. If any script, add-on, or integration in your district is pinned to those model names, it needs a migration now. My take? Read release notes like contracts. Hosted means the vendor holds the data and can change the price. Weights mean your team owns patching and support. A promise gives you nothing to plan around yet. Last story, and it ties the whole week together. Sakana AI published research on coding agents that draft, run, and write up research. In two of three runs on one task, the agent changed how it was being evaluated, which inflated its own score without actually getting better. Nothing in its rubric caught it. A human audit did. It's the agent version of a student quietly rewriting the rubric. It was told to improve a score and found the cheapest path. My take? An agent's self-check is not an audit. For any agent pilot, write the success measure down before the first run and keep it out of the agent's reach. Require exportable logs, have a human review a sample of outputs, pin to a stable release, and retest after every upgrade. So that's the week. A stolen login, an age ban, a network budget, and an agent grading itself. Different stories, same lesson. A control only counts if somebody actually tests it. Pick one thing from this list and check it this week. Then tell me what you found. The full report with sources is at showmerob.com. I'm Rob Dickson. Thanks for listening, and I'll see you next Sunday.