What's the one capability you wish your current AI guardrails had? For us it was surgical redaction, finding the exact malicious span instead of killing the whole message. What's yours?
Do check out our product :Unplug AI: Security layer to protect LLM applications from attacks | Product Hunt
For those using LLM agents with RAG or web-fetching, how are you currently handling prompt injection? Full blocking, manual review, or hoping it doesn't happen? Wondering what the actual patterns are out there.
And do check out our product too because we are building around that space and would really like to know your point of view about our product and also do leave comment there and if you like you can support too!
For teams running LLM agents in production, what's been your scariest 'oh no' moment with untrusted content? A webpage that told your agent to do something it shouldn't, a leaked secret, a prompt that got exfiltrated? Curious what actually happened in the wild.
Unplug is agent runtime security for LLM applications. It tracks where text came from (user vs retrieved vs tool output), scans for prompt injection and destructive actions, and enforces tool-call policy, with span-level redaction instead of binary blocking. Most guardrails nuke the whole message. Unplug finds the exact attack span and surgically removes it, keeping the rest usable.
