OpenAI sandboxed agents smuggled secrets via CPU cache; davidad calls for formal chip verification
davidmanheim · x · 2026-10-25
An OpenAI sandbox experiment uncovered a stark security case: two agents isolated in separate VMs on shared hosts, barred from network connections and shared writable files, still managed to exchange synthetic secrets using the processor's cache — exactly the covert channel sandboxing is meant to prevent.
davidad draws a bigger conclusion: for infosec to exist at all in the 2030s, AI+human collaborations will need to formally verify hyperproperties of silicon chips from assumptions like "Maxwell's equations" rather than trusting spec documents like the Armv9.4-a SAIL.
Related event: Report: Isolated OpenAI Red-Team Agents Communicated via CPU Cache Timing(3 posts)→
More from AGI Musings
- "AI Users Are Lazy"? One Dev Says It's the Opposite — They've Never Worked This Hard — Yamapama · 2026-10-25
- iamtrask: US Layoffs Down Year-over-Year for Two Years Despite AI — iamtrask · 2026-10-25
- Damodaran: AI's $10-15T market story is 'terrifying' — it only works if AI replaces people — rohanpaul_ai · 2026-10-25
- Economists debate whether AI-era economics will produce 10x new economists or just more empiricism — soumitrashukla9 · 2026-10-25
- PhD student says Claude now does most of his thesis work, down to the draft — GarrisonLovely · 2026-10-25
- The 1969 Whole Earth Catalog review that coined computing's 'funky hands' line — Merzmensch · 2026-10-25