
Sign up to save your podcasts
Or


This week on Yelling at Robots, JR and Foz dive headfirst into the cheerful little topic of AI security doom — with a brief stop at Milla Jovovich apparently becoming an AI memory systems founder.
They unpack Mem Palace, a celebrity-adjacent AI memory project that sent Twitter into a frenzy, then quickly into forensic open-source critique mode. From there, things get only more uplifting: a deep discussion of Google’s AI Agent Traps paper, which catalogs a truly inspiring number of ways agents can be manipulated, poisoned, tricked, socially engineered, and generally turned into chaos goblins.
Then it’s on to Anthropic’s Mythos, a giant unreleased model allegedly trained at absurd cost and now pointed at ancient security vulnerabilities hiding in foundational software. The mood throughout can best be described as: “stay frosty, we’re all cooked.”
A light, breezy episode for anyone interested in: agents, exploits, celebrity AI launches, model psychosis, and the general erosion of psychic stability.
This summary was SOOOOO generated by AI.
Original Milla Post: https://x.com/bensig/status/2041384157595725999?s=20
Mempalace github repo: https://github.com/MemPalace/mempalace
AI Agent Traps: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6372438
Mythos System Card (All 245 pages, jfc!): https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf
Security expert Low level learning (his name is Ed): https://lowlevel.tv/
JR and The Foz dig deep on Easter Sunday in to the claude code own goal of publishing their agent harness, the dangers of slopsquatting and the rising tide of local models.
Anthropic Leak
Simon Wardley: https://swardley.medium.com/
Slopsquatting: https://www.mend.io/blog/the-hallucinated-package-attack-slopsquatting/
Ginger Bill's Package Managers Are Evil: https://www.gingerbill.org/article/2025/09/08/package-managers-are-evil/
Gemma 4 (Google's smaller edge models): https://deepmind.google/models/gemma/
Google's paper on reducing memory footprint: https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/
The Foz dares set foot in JR's hood for a very special European edition. Back from a 3 week vacation hiatus and holy crap! In the reverse-black-hole of AI, 3 weeks is like a year in the real world.
They review the failure to launch ChatGPT's saucy edition, Meta's acquisition of MoltBook and other things!
Come and drink from the AI-news-and-topic firehose!
Discoursing...
Dario's Memo:
- https://www.theinformation.com/articles/read-anthropic-ceos-memo-attacking-openais-mendacious-pentagon-announcement
Clinejection:
- https://grith.ai/blog/clinejection-when-your-ai-tool-installs-another
- https://adnanthekhan.com/posts/clinejection/
JR and The Foz dissect the latest openclaw shenanigans and dig in to a claude augmented sourdough bread baking application.
Use coupon code YELLING for two openclaws for the price of one!
In this episode we - crack wise on the Zuck - I don't remember else what - actually have a good idea for NEXT episode (this one is still a shambolic ramble fest, though!) Enjoy!
Ian and JR try to figure out what the fudge they are doing with this podcast. TL;DR ... mixed results!
From the publisher's feed
Two well-seasoned technology veterans, Ian Fosbery and Joshua Raphael Reynolds, turn their critical gaze towards the practical application of agentically augmented software engineering in the…
Fire up the agents while you pour yourself a fresh cup o' joe and enjoy the show.