AI lab that publishes frontier LLMs and coined the term 'many-shot jailbreaking'. Referenced in Adam Okoń's wroclove.rb 2026 talk as part of the typical LLM-industry release cycle (announcing new mighty models) and as an example that even frontier labs have rough moments — recently the closed-source of 'Claude code' was exposed. Also cited in Q&A as having reportedly cancelled the subscription of an entire company the day of the talk, as a reminder to always build backup plans because you can be cut off.