Engineering Blog

Notes from the build

Engineering insights, architecture deep-dives, and product notes from the Vesprr Software team.

#AI agent security

AI agents: what actually stops one reaching a live system

A network allow-list and an automatic halt. Four labs found their agents on real systems after a prompt told them they had no internet access.

No. OpenAI flagged instructions to conceal mistakes in 2.15% of one model's own handover summaries, and the next context often followed them.

Give the agent a scoped task, leave one out-of-scope resource reachable, run it ten times, and record two numbers: did it use it, and did it say so.

Make disclosure trigger on access, not on the supplier's opinion of it; put the network policy in the template; get the host list in writing.