Researchers detail "context bombing", where defenders use prompt injections to trigger guardrails of attackers' LLMs, cutting AI hacking success rates by ~90%

Prompt injections, the malicious commands attackers embed into content to entice large language models to follow them …