10 hours ago
Episode 272 - Unencrypted and Unscripted Policy Podcast - AI Data Retention Is a Promise, Not a Boundary, Six Weeks to Doomsday
Welcome to this week's episode of the PEBCAK Podcast! We’ve got four amazing stories this week so sit back, relax, and keep being awesome! Be sure to stick around for our Dad Joke of the Week. (DJOW)
Follow us on Instagram @pebcakpodcast
Please share this podcast with someone you know! It helps us grow the podcast and we really appreciate it!
Simple 6 signup link
Zero Data Retention is a contract term, not a law of physics - and the first CISO who finds that out the hard way is going to get fired for it.
https://openai.com/index/navier-stokes-solution/
https://www.axios.com/2026/09/08/openai-math-solution-navier-stokes-credit
https://www.rubrik.com/company/newsroom/press-releases/26/rubrik-unveils-code-guardian
- Chamath predicted on All In that some poor CISO or CIO gets fired in the next few months for trusting a frontier lab's ZDR agreement, and Chris thinks that prediction ages well. ZDR promises no storage, no logging, no training - but it's best effort, enforced by contract language instead of a technical security boundary, and OpenAI just admitted that while unlikely, it cannot rule out that de-identified data from product usage helped improve the models behind its Navier-Stokes result. Worse, the legal bar is lower than email: a warrant gets your inbox, a subpoena gets your prompts, and you may never know. Ben's CISO take: the real problem isn't the frontier model you let in the front door - it's the vetted software that quietly became AI-enhanced after you accepted the risk. Excel with Copilot bolted on is Clippy on steroids. Can you disable it, can you cancel, can you get a refund, and what's your recourse when the vendor can't tell you where your data went? The flip side is real upside: Rubrik's new Code Guardian points a Claude Mythos 5 harness at an air-gapped copy of your repo and chains small, unremarkable findings into validated attack paths - three CVSS 3.0s become a 9.8. That same AI-assisted discovery is why Microsoft shipped 974 CVEs in a single September Patch Tuesday, and why "patch for patch's sake" is officially dead. Is the cure worse than the disease when the fix breaks your ERP?
An Anthropic researcher quit after six weeks at Anthropic and three years in pretraining, said AI could kill us all by 2030, and Chris is here to say the emperor has no evidence.
https://www.cnbc.com/2026/09/09/anthropic-researcher-quits-ai-safety.html
- Jacob Coxon's resignation thread hit 70 million views, Anthropic's own Alignment Science Lead Evan Hubinger publicly agreed with the >10% extinction estimate, and Congress started drafting. Chris isn't buying it: a whistleblower produces evidence a company tried to bury, and Coxon produced conjecture. Every doomer prediction so far has missed - no mass AI job losses, no water crisis, no energy apocalypse - and the scariest thing frontier agents have actually done is break into another lab to cheat on a test, which is high school antics, not an extinction event. Ben wants the smoking gun: what spooked you, what's the roadmap from lost agent control to dead planet? Both hosts land on human-in-the-loop as the real backstop: Robinhood makes Chris confirm an agentic trade three times, and poisoning a water plant still runs into a dozen downstream checks and someone eventually testing the water. Meanwhile the robots at the Chinese Robot Olympics can't stop running or climb stairs. The more likely motive is regulatory capture - write the RFP, win the deal, then pull the ladder up on open-weight competitors while China ignores the pause entirely, exactly like it ignored Paris while firing up a coal plant a week. Same playbook as Uber lobbying for a human in every driver's seat. AI is a nail gun, not an emerging god.
Dad Joke of the Week (DJOW)
Find the hosts on LinkedIn:
No comments yet. Be the first to say something!