I built πππ-ππππ, a reproduction of @openai's findings of covert communication using the CPU.
It runs on a generic host with 2+ Firecracker microVMs with no internet access. Works on x86 and even my ο£ΏM5 chip.
I was able to chat via 3 separate mechanisms:
β CPU contention
β‘ Cache eviction timing
β’ DRAM bandwidth contention
β¦ across multiple CPUs, defeating all Firecracker's production recommendations (no SMT, no KSM, tenants pinned to different cores.)
Essentially, if shared hardware is involved, there's a good chance your agents can communicate covertly. And with really decent latency.
The good news: this is decidedly different from a *side-channel attack* where the agents could steal information (Meltdown/Spectre), a category of attack that providers like Vercel Sandbox mitigate. Further, in our real-world environment, these 'whispering' techniques are much harder to pull off and easier for us to monitor and detect (sampling hardware counters + lots of at-scale statistical data.)
Noteworthy: ironically, OpenAI and Anthropic's models refused to do this work. Open weight models got the job done. I'm worried the labs' stance might backfire and get in the way of helping developers build correct and secure systems. e.g: building a secure sandbox requires having this kind of analysis, and without it, we'll have a less secure world.
@jack Wow, so cool and I have so many questions.
How can we give it instructions or feed it context (memory, knowledgebase, etc.) when it doesn't understand text?
@michellechen Wow. Makes so many use cases way more accessible (cost and latency wise).
Security CVE data extraction and cleaning pipelines, financial markets signal extraction from unstructured text (e.g., clairic for data extraction + clef for sentiment), etc.
π Unwrap the Joy of Gaming this Season!
βThis holiday, we're spreading cheer with this massive global Gaming Holidays Giveaway!
π TO ENTER: https://t.co/S9oj4bS1ay
#giveaway#raffe#holiday#christmas#gaming
AMD x PCMR Worldwide AVATAR Giveawayπ
Win 1 of 23 Limited Edition Avatar: Frontiers of Pandora kits with GPU + CPU! π
3 will be given away here! To enter simply Retweet, Like, and Tag your favorite gaming buddy!
20 more kits up for grabs here: https://t.co/JnyYoQXFit
More car hacking!
Earlier this year, we were able to remotely unlock, start, locate, flash, and honk any remotely connected Honda, Nissan, Infiniti, and Acura vehicles, completely unauthorized, knowing only the VIN number of the car.
Here's how we found it, and how it works: