Thank you very much Shravan!

Strangely, the above discussion fits well the writing grants objective function.

AI welfare is a thing: www.transformernews.ai/p/anthropic-...

#security #privacy #AI #LLM

This is the first evidence that a challenging attack on AI, i.e., prompt injections, which the community cannot eliminate (as of now), can be turned into something constructive and positive for the security community

(2) once the LLM-agent falls for the trap, then we deploy a prompt injection that can either eat up the resources of the attacking LLM (by luring it to keep looking in an infinite filesystem) or even persuade the LLM-agent to run a reverse shell so the defender can take over the attacker's machine!

This year, we saw the first automated cyberattacks in the wild where LLMs dictated the strategy. We built a proactive defense system to mitigate this threat. Our idea was to (1) exploit the greedy nature of LLMs and lure them to a decoy service that looked vulnerable but unrelated to the real system

This year at IEEE S&P we tried a shorter talk but with a mandatory poster slot immediately after the session (in face in the same room). Worked surprising well! Talk is seen as a teaser for a more elaborate discussion during the poster. Also posters were not crowded cause it was per session.