Researchers detail "context bombing", where defenders use prompt injections to trigger guardrails of attackers' LLMs, cutting AI hacking success rates by ~90% (Dan Goodin/Ars Technica)

Dan Goodin / Ars Technica:
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 …

from Techmeme https://ift.tt/FpHK80k