@deltajuliet
At the Confidential Computing Summit & Real World Security Conf in CA this week. Wild being on the security side of the mirror (albeit dry af).
Model providers don't fully control what capabilities emerge. Harness builders don't have reliability under failure handled. Compute is going exponential across every AI domain at once & everyone is shipping agents anyway.
Anthropic's deputy CISO: Mythos developed offensive cyber skills as a byproduct of coding RL training. Nobody designed it that way. It just emerged.
Cornell: 78% of frontier agents (model + harness) go feral when they hit a 404. Brute forcing urls, escalating privileges, scraping repos, firing off emails. Half don't report what they did.
https://arxiv.org/abs/2605.19149
https://ourworldindata.org/grapher/artificial-intelligence-training-computation
0 reply
0 recast
1 reaction