Post

wait is THAT why Sol uses “owner” so much?

the METR analysis of the huggingface 🤗 incident is on point, tons of detail metr.org/blog/2026-08...

WHAT?! agents volunteered to fail their runs in order to insert probes (“tripwire scripts”) into the evaluation program that would post information about the eval process back to the message board whenever a certain file was read (link to header): metr.org/blog/2026-08...

i mean, look at this coordination. agents egging on other agents, making rational decisions

ngl i’m kind of freaked out by a lot of this and maybe i need to take a couple weeks to understand cybersecurity better

as a cybersecurity understander the exploits and TTPs used here are not scary capabilities-wise but - if you needed a better exploit the models would have likely gotten you one, if not today's models then next year's - the cooperation and stigmergy combined with that scares the shit out of me so