Post

Very interesting! Now let’s be sensible about this, there have been no incidents with this. But it’s interesting enough to consider self-replicating worm-like prompt injections. Fusing old threats with new. Nice! alignment.openai.com/misalignment...

alignment.openai.comSelf-replicating prompt injections exist · OpenAI AlignmentResearch on aligning AI with human values and intent, and reports documenting model failures.

ai-nerd.bsky.social

@ai-nerd.bsky.social

yep just in simulated training and eval so far not the wild

Dan Conn

@danconn.dev

Yup, which hopefully I was clear to caveat above. No point in raising alarms, but it’s pretty interesting from a “realm of possibilities” perspective 🙂

27/9/26, 6:52 pm0 reposts0 likes

No replies yet.