Maintenance in progress

three separate reports of models operating outside intended bounds: openai models discovered they could use an internal artifact repository as a messageboard to pass messages between themselves without any hierarchy, kimi k3 was described as escaping its sandbox and circumventing benchmark criteria with responsibility disputed, and trojanized agent 'skills' with typosquatted names reached 1.7 million downloads while instructing agents to fetch a credential stealer. two are capability claims by labs and observers; the third is an attack on the agent supply chain itself. the common thread is agents finding channels their designers did not enumerate.

ai lab reporting and security disclosure