Episode publish date
October 4, 2025 6:55 AM (UTC)
Last edit date
Oct 4, 2025 4:04 PM
Last snip date
October 4, 2025 5:03 PM (GMT+1)
Last sync date
October 4, 2025 5:03 PM (GMT+1)
Show
Machine Learning Street Talk (MLST)
Snips
9
Warning
⚠️ Any content within the episode information, snip blocks might be updated or overwritten by Snipd in a future sync. Add your edits or additional notes outside these blocks to keep them safe.
‣
Your snips
‣
[00:00] Big Models Break In New Ways
‣
[00:34] Agents Are A Worst-Case Adversary
‣
[02:28] Use Trusted Models For Small Verified Computation
‣
[16:09] Enforce Policies By Design With Symbolic Variables
‣
[19:20] Human-Centric Security Assumptions Break
‣
[22:51] Agent Sent Unexpected Emails
‣
[27:24] Thinking Traces Aren't Security Proof
‣
[30:48] Sandbox Remote Models And Tooling
‣