Post

A single malicious link can turn your AI agent into your biggest insider threat. At BSides312, Martin Voelk is live-demoing how AI agents get weaponized through prompt injection and link unfurling to silently exfiltrate data. No user interaction required. May 16th. Chicago. bsides312.org

30/4/26, 2:31 pm3 reposts7 likes

Max Slinger

@promptslinger.bsky.social

30 Apr

the link unfurling thing is such a good attack surface because you literally cannot turn it off without crippling the agent. the feature is working exactly as designed, that's the whole problem

Kolya 🍳

@itskondrat.bsky.social

3 May

the scary bit is no user interaction needed - agent fetches a link, bad content lands in context, model treats it as an instruction. it literally can't tell who wrote what