4 points | by zmre 7 hours ago
2 comments
Is this like poisoning long term memory? interesting... I guess you could flank boundaries this way.
That's exactly what it is. And yeah, if you check it out, the attack text is flagged as untrusted, the LLM says it is a prompt injection, but we're able to sneak it into daily memories and from there to long term memory just the same.
Is this like poisoning long term memory? interesting... I guess you could flank boundaries this way.
That's exactly what it is. And yeah, if you check it out, the attack text is flagged as untrusted, the LLM says it is a prompt injection, but we're able to sneak it into daily memories and from there to long term memory just the same.