Bypassing the Safety Training of Open-Source LLMs with Priming Attacks