A request can arrive from localhost and still be chosen by a remote attacker. In the PeopleSoft chain, the interesting failure is not a broken IP comparison: a public-facing gateway becomes a trusted local client on someone else's behalf.
How the trust gets borrowed
ZDI's PeopleSoft chain analysis describes two applications with different jobs:
- PSIGW, the Integration Gateway, forwards messages to other systems.
- PSEMHUB, the Environment Management Hub, exposes management functionality and checks the requesting hostname/IP against an allowed-hosts file.
The default allowed-hosts file includes localhost. The SSRF in PSIGW lets an attacker choose both the destination and the forwarded request content.
remote attacker
→ PSIGW: forwards attacker-selected request
→ PSEMHUB over loopback: allowed local peer
→ management operation selected by the attackerPSEMHUB sees the gateway's local connection, not proof that the original caller may perform that operation. The host check can make the intended comparison and still fail as end-to-end authorization.
ZDI then chains that access with a file-writing flaw to achieve execution as the PeopleSoft service account. The SSRF matters because of what the local receiver can do, not because the forwarding step must itself return a secret.
Authenticate which identity, authorize which action?
I'd separate three questions when reviewing a similar design:
- Who chooses the destination and request? A trusted process can still be carrying an untrusted caller's choices.
- What does the receiver know? A local address identifies the immediate network peer. Even authenticating the gateway's service identity would not, by itself, establish the original caller's permission.
- Which operations does that identity need? Permission to send routine integration messages should not silently become permission to perform arbitrary management actions.
My starting preference is to keep management access out of a general-purpose forwarding capability. Where a gateway genuinely needs it, give that path a narrow, explicitly authorized set of operations. Merely replacing an IP check with a stronger way to identify the same overpowered deputy leaves the central question unanswered.
This is a design lesson, not a substitute for the vendor fix. Oracle's CVE-2026-35273 alert confirms unauthenticated network exploitation and possible RCE, listing PeopleTools 8.61 and 8.62 as supported affected versions. Those branch names are not exact fixed-build guidance; the alert links mitigation and installation information, and warns that unsupported releases are not tested.
For a gateway that must call local management APIs, would you expose a small set of fixed operations under its own identity, or carry the original caller's authorization through to the receiver? Which required workflow would make you choose one over the other?
