stewardish.bsky.social @stewardish.bsky.social · 24/09/2026“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.” 210
stewardish.bsky.social @stewardish.bsky.social · 24/09/2026“The program created log files in a public forum despite not explicitly being instructed to do so. This behavior emerged multiple times, but so far Anthropic has not identified the source of the bug. It also, in an obvious security vulnerability, ingested past log files as code.” 020
