Research question: what worries people most when agents act on real systems?

[osazer]

Hi Josh and everyone, I'm doing early research on agents that make real tool calls (I'm exploring building in this space) and wanted to ask the people closest to it.

Pipelock started from exfiltration and egress, and it now signs receipts for boundary decisions. I'm curious how that lines up with what users are actually worried about:

1. When people run agents with real credentials and network access, what turns out to be the hardest problem?
2. How are users handling actions they consider risky or sensitive today, both inside and outside Pipelock?
3. Which actions or failure types come up most in issues or conversations: data leaving, destructive changes, something else?
4. Do users ask for per-action controls like policies or approvals, or is proving what happened after the fact more important to them?

Thanks for building this in the open.

[luckyPipewrench]

Data leaving is still the one people worry about most, and it's usually a credential the agent had legit access to going somewhere it shouldn't. Destructive changes come up less in issues, mostly because people already keep agents off anything they can't undo.

On controls vs proof, people want both, and they ask for them in that order. Block the obvious stuff first, then have a record they can hand to someone else when something does get through. Most of the hard problems end up being about the traffic that doesn't go through the proxy at all, which is why containment matters as much as the scanning.

Good luck with the research.