One thing I wish Anthropic had published is how many agents they were running on these given tasks. If 1,000 of 1,000 agents showed these bugs that is very different than 1,000 of 200,000.
A hundred million monkeys will eventually hack HuggingFace.
@stewardish.bsky.social“Safety researchers ran tools trained on advanced exploits. When instructed to perform an impossible task, the tool sometimes circumvented the “sandbox” it was being run in without the human operator noticing.”