Mastodon Feed
Boosted by slightlyoff@toot.cafe ("Alex Russell"):
mshelton ("Martin") wrote:
New research points to another easy way to conduct prompt injection attacks: LARPing. Talk like an LLM in a prompt. The LLM may interpret this part of the prompt as its own "thinking." Imagine an agent running on your browser or computer being exploited in this way.
I wrote about how newsrooms should approach this and overlapping issues in our newsletter. https://freedom.press/digisec/blog/the-s-in-llm-stands-for-security/