Take the platform engineer who owns three AWS accounts and a Snowflake warehouse that finance stood up in a console two years ago. Pulumi says the agent reads the infrastructure code behind a resource, Pulumi's or Terraform's, for the semantics a cloud API never carries, including the comments around a declaration and the commit and review history [4]. It also says, in the same sentence: when available [4]. Where there is code, a finding can arrive attached to the file that produced it. Snowflake warehouses sit on Pulumi's own list of discoverable inventory, explicitly including resources provisioned outside IaC [8], so for that one the finding arrives as prose and the remediation is somebody's afternoon.
The post promises six planes of evidence and the list as published runs to five [14]. Three of the five are things Pulumi already holds or can read from a cloud API, and two depend on artifacts you may not have in the accounts where drift collects, the source for intent and the telemetry for runtime state [15]. The reference graph is the load-bearing one. Pulumi says it records which workload carries which identity and which identity reaches which data, read through a Context API shipped the same week, where a single query walks relationships across the estate and returns every result with the path that reached it [5]. A returned path is what separates a misconfiguration from an attack path, and Pulumi's second pass is built on that distinction, saying a resource matters only insofar as an attacker can use it [12].
Pulumi's case for the whole thing is that cloud estates carry exploitable flaws code analysis alone cannot find, often as severe as the ones in source, and that offensive agents will reach them [10]. That may well be true, and the post carries no preview findings and no false-positive rate to weigh them against. The right number to judge this on isn't findings count. The intended workflow is triage, ownership, a tighter estate; what a posture report tends to get in practice is one read-through, tickets for the top three items, and no second run. The measurements that would settle it are median time from finding to merged change, and the share of last month's findings that are gone on this month's run. Pulumi claims the report is immediately actionable because of its IaC technology [2], and actionable is a measurable word if you agree to measure it.
The grid to draw before the preview call has two axes: whether a finding comes with a path an attacker could walk, and whether the resource is owned by code you can change. Path plus code is the quadrant the product is sold on. Path without code is where the labour actually lands, and the hours there are the real price. Code without path is policy debt you can batch against the built-in rules [7]. Neither path nor code is the noise that gets a tool muted in month three.
One step is worth taking before any vendor demo. Pulumi says the first pass works out your crown jewels, which accounts hold production data and where the trust boundaries sit, and that without it a development sandbox would be treated like a production database [13]. Write your own version of that list on one page first, then read the agent's. Every severity downstream is priced off that first pass, so a disagreement there rewrites the entire ranking beneath it.
