Add libssh2 CVE-2026-55200 benign PoC

This commit is contained in:
ashton
2026-06-23 04:25:35 -05:00
parent 37ff6d3233
commit a638a1cf5e
4 changed files with 600 additions and 1 deletions

View File

@@ -0,0 +1,164 @@
# libssh2 CVE-2026-55200 benign PoC
Benign local proof of concept for CVE-2026-55200, an unchecked SSH `packet_length` condition in libssh2's `ssh2_transport_read()` transport parser path.
Research status: verified local arithmetic and state model.
## Summary
libssh2 through 1.11.1 accepted an attacker-controlled SSH packet length in one full-packet decryption path without first enforcing the RFC-sized libssh2 packet maximum.
The vulnerable shape is:
```text
total_num = 4
packet_length = decoded SSH packet_length field
reject only if packet_length < 1
total_num += packet_length + mac_len + auth_len
reject if total_num > 35000 or total_num == 0
allocate total_num bytes
```
On a build where `size_t` is 32-bit, a packet length of `0xffffffff` with `auth_len=16` wraps the computed allocation size to 19 bytes:
```text
4 + 0xffffffff + 0 + 16 == 19 modulo 2^32
```
The original `packet_length` remains `0xffffffff`. Later full-packet processing can still use packet-length-derived sizes, including a `packet_length - 1` style length, after the allocation decision has already been made.
The upstream fix rejects `packet_length > LIBSSH2_PACKET_MAXPAYLOAD` before the addition.
This PoC does not generate SSH traffic, does not attempt remote exploitation, and does not perform an out-of-bounds write. It is a standalone verifier for the arithmetic and decision-state transition.
## Files
- `poc/cve_2026_55200_probe.c` - standalone C11 benign verifier with no source comments
- `evidence/2026-06-23-local-harness-output.txt` - local build and replay evidence
## Affected Source Path
The affected path is in:
```text
src/transport.c:ssh2_transport_read()
```
The fix is upstream commit:
```text
97acf3dfda80c91c3a8c9f2372546301d4a1a7a8
```
The relevant fix adds a `packet_length > LIBSSH2_PACKET_MAXPAYLOAD` guard before the vulnerable addition in the full-packet path.
## Build and Run
Linux, macOS, WSL, or MinGW:
```bash
gcc -std=c11 -Wall -Wextra -O0 -g -o cve_2026_55200_probe poc/cve_2026_55200_probe.c
./cve_2026_55200_probe
```
Windows PowerShell with MinGW:
```powershell
gcc -std=c11 -Wall -Wextra -O0 -g -o cve_2026_55200_probe.exe .\poc\cve_2026_55200_probe.c
.\cve_2026_55200_probe.exe
```
No arguments defaults to the benign proof mode. The same mode can be selected explicitly:
```bash
./cve_2026_55200_probe --benign
```
Additional diagnostic modes:
```bash
./cve_2026_55200_probe --native
./cve_2026_55200_probe --check
```
The input values can be overridden:
```bash
./cve_2026_55200_probe --packet-length 0xffffffff --mac-len 0 --auth-len 16
```
## Expected Output
On a 64-bit build, the default benign run should report `result=PASS`:
```text
benign CVE-2026-55200 proof
build_size_t_bytes=8
build_size_t_bits=64
packet_length=0xffffffff (4294967295)
mac_len=0
auth_len=16
mathematical_total=4294967315
vulnerable32_decision=accepted
vulnerable32_total=19
vulnerable32_allocation=19
fullpacket_style_length=4294967294
allocation_gap=4294967275
fixed32_decision=rejected: out of boundary
native_unpatched_decision=rejected: out of boundary
native_unpatched_total=4294967315
native_note=64-bit native arithmetic rejects this default input; modeled 32-bit arithmetic remains vulnerable
result=PASS
```
The `--native` mode shows what the current binary's actual `size_t` arithmetic does. On a 64-bit build, native arithmetic rejects the default input because the computed total does not wrap before the existing `total_num > 35000` check:
```text
native-size_t check
build_size_t_bytes=8
build_size_t_bits=64
unpatched_decision=rejected: out of boundary
unpatched_total=4294967315
unpatched_allocation=0
fixed_decision=rejected: out of boundary
fixed_total=0
fixed_allocation=0
```
## Mechanics
The verifier computes three decisions for the same input:
- `vulnerable32`: the vulnerable arithmetic using a 32-bit `size_t` stand-in
- `fixed32`: the patched 32-bit decision with the packet-length maximum check
- `native_unpatched`: the unpatched decision using the current binary's real `size_t`
For the default input, the modeled 32-bit vulnerable path accepts the packet and derives a 19-byte allocation. The fixed model rejects before allocation. A native 64-bit build also rejects the default input, but that does not invalidate the 32-bit vulnerable state; it shows the architecture dependency explicitly.
The proof condition is:
```text
vulnerable32_decision == accepted
vulnerable32_allocation == 19
packet_length > 35000
fullpacket_style_length > vulnerable32_allocation
fixed32_decision == rejected
```
## Why This Is Benign
The program does not connect to any host, bind any socket, generate any SSH packet, or copy past an allocation. It only models the arithmetic and branch decisions needed to distinguish the vulnerable and fixed behavior.
The `allocation_gap` value is a diagnostic showing how far a later packet-length-derived operation would exceed the modeled allocation if such an operation were performed.
## References
- CVE record: `https://www.cve.org/CVERecord?id=CVE-2026-55200`
- NVD entry: `https://nvd.nist.gov/vuln/detail/CVE-2026-55200`
- Upstream fix: `https://github.com/libssh2/libssh2/commit/97acf3dfda80c91c3a8c9f2372546301d4a1a7a8`
- Upstream PR: `https://github.com/libssh2/libssh2/pull/2052`
- VulnCheck advisory: `https://www.vulncheck.com/advisories/libssh2-out-of-bounds-write-via-unchecked-packet-length-in-transport-c`
## Responsible Use
Run this PoC only as a local research and regression-verification harness. It is intentionally scoped to a non-network arithmetic model.