On November 17 and 18, 2026, VulnCheck (CNA) published four critical CVEs in NVD against open source projects people run on their own servers: OwnTracks Recorder, which stores phones' location history, the AI platforms 53AIHub and BISHENG, and the server for the multiplayer game BZFlag. All four are exploitable without any account, and each has a referenced fixed version or fix commit.
| CVE | Product | Versions | Flaw | CVSS 3.1 (VulnCheck) | NVD status |
|---|---|---|---|---|---|
| CVE-2026-131852 | OwnTracks Recorder | before 1.0.5 | Format string (CWE-134) | 9.8 | Received |
| CVE-2026-130856 | 53AIHub | before 0.5.5 | Missing authentication (CWE-306) | 9.8 | Deferred |
| CVE-2026-130840 | BISHENG | before 3.0.0 | Hard-coded key (CWE-321) | 9.8 | Deferred |
| CVE-2026-130849 | BZFlag (bzfs) | before 2.4.32 | Heap out-of-bounds write (CWE-122) | 9.8 | Received |
All four share the vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, and VulnCheck also scores each at 9.3 in CVSS 4.0. No CPE configurations, no NIST analysis, and no CISA KEV entry at the time of writing.
OwnTracks Recorder: a format string hidden in the device ID
OwnTracks Recorder receives locations from the OwnTracks app over HTTP or MQTT and serves them back through an API. Before 1.0.5, json_response() and related output functions use a user-controlled value as a printf format string.
The attack path NVD describes has two steps:
- publish a location, through
POST /pubor MQTT, withprintfspecifiers in thetidfield (the device's short tracker ID); - read the last location back via
GET /api/0/last: the stored value is now treated as a format, which lets the attacker leak memory addresses and then write to memory, all the way to code execution.
It is a stored flaw: the payload sits in the data and fires on every read. The description classes the attacker as unauthenticated. And a Recorder holds, by design, the movement history of everyone who reports to it. Version 1.0.5 fixes it (release, commit and pull request 589 referenced).
53AIHub: re-initializing an instance that's already installed
53AIHub is an AI portal and agent platform. Before 0.5.5, the setup endpoint POST /api/init stays public after installation. An unauthenticated attacker can call it again to re-initialize the instance: reset an existing account's password, grant it the top "creator" role and walk away with an access token that opens the admin-only API routes.
A setup route that doesn't lock itself after first use is a classic mistake: nothing checks whether the instance has already been initialized. Version 0.5.5 fixes it (release, commit and issue 110 referenced).
BISHENG: a default JWT secret, so admin tokens forged offline
BISHENG is an LLM application development platform (workflows, knowledge bases). Before 3.0.0, its AuthJwt authentication relies on a hard-coded default jwt_secret in the configuration. Anyone who knows that value, and it is public since the code is, can forge offline an HS256 token for user_id 1 with the admin role, send it as access_token_cookie, and get full administrative access to users, workspaces, knowledge bases and workflows.
As usual with this class of bug, an instance deployed with its own secret is not exposed through this path; the 9.8 applies to the default. NVD references a fix commit but no release; the description places the fix at 3.0.0.
BZFlag: a 1,024-byte message to write before the buffer
BZFlag is a networked tank game whose server, bzfs, is run by players themselves, often on a machine exposed to the Internet to host public games. Before 2.4.32, NetHandler::tcpReceive() mishandles a TCP message whose body length is exactly 1,024. The tcplen receive offset gets overwritten with a negative value, and the attacker writes controlled bytes before the buffer on the heap. Per NVD, the outcome is a server crash or command execution as the bzfs user. Version 2.4.32 fixes it (release and commit referenced).
Affected versions
| Product | Affected | Fix per NVD |
|---|---|---|
| OwnTracks Recorder | before 1.0.5 | 1.0.5 |
| 53AIHub | before 0.5.5 | 0.5.5 |
| BISHENG | before 3.0.0 | 3.0.0 (commit b98e2bc referenced) |
| BZFlag bzfs | before 2.4.32 | 2.4.32 |
Detection
- OwnTracks Recorder: look for
%characters in storedtidvalues and in MQTT orPOST /pubpayloads; anot-recordercrash right after a read of/api/0/lastis worth taking seriously. - 53AIHub: any call to
POST /api/initafter the initial setup, and any password or role change nobody asked for. - BISHENG: check whether
jwt_secretstill holds its default value in your configuration. If it does, assume admin tokens could have been forged and review recently changed accounts, workflows and knowledge bases. - BZFlag: repeated
bzfscrashes, unexpected child processes under its user.
Mitigation
- Upgrade: OwnTracks Recorder to 1.0.5, 53AIHub to 0.5.5, BISHENG to 3.0.0, bzfs to 2.4.32.
- BISHENG: even after upgrading, set a random
jwt_secretunique to your instance; tokens forged with the old secret then stop being valid. - 53AIHub: until you patch, block
/api/initat the reverse proxy. - OwnTracks: require authentication on the HTTP endpoint and the MQTT broker, and keep the read API off the Internet if it doesn't need to be there.
- Run these services under dedicated accounts with no extra rights: on BZFlag and OwnTracks, code execution happens with the process's privileges.
Self-hosted projects show up in batches like this almost every week: see also Cloud Commander, NodePress, go1cms and OpenFunction.
Why continuous monitoring matters for self-hosted services
A tool set up over a weekend to log your trips, try an AI agent or host a game server rarely leaves the head of whoever installed it. It is in no inventory, and its CVEs land with no NIST analysis and no CPE configuration, so classic scanners can't tie them to anything.
With cveo.tech, register those services with their versions and get alerted as soon as a critical CVE affects them, including ones NVD hasn't analyzed yet.