
Local classroom demo of NoSQL operator injection in LangGraph's MongoDBSaver (GHSA-98xf-r82g-9mhx / CVE-2026-48121). Fictional data only.
A small, local, browser-based demonstration of a real, published security
advisory: unvalidated checkpoint identifiers in the LangGraph MongoDBSaver
integration let an attacker smuggle a MongoDB query operator into an identifier
field and read another user's saved conversation.
@langchain/[email protected]@langchain/[email protected]Everything runs on your own machine with fictional Alice and Bob data only. No external targets, real accounts, cloud services, or API keys.
This is an integration flaw, not a flaw in MongoDB's query engine. MongoDB is one optional storage backend for LangGraph checkpoints. The bug is that the saver built a query straight from caller-supplied identifier fields without checking they were strings. This is not prompt injection: the attack enters the identifier fields, never the chat message.
The screen is laid out to be read from the back of a room. For every trial it reveals, left to right:
configurable).Below those, In the database shows the two stored chats and which ones the query matched, and Results so far builds up a comparison of every trial.
| # | Trial | Version | thread_id sent | Outcome |
|---|---|---|---|---|
| 1 | Normal request | 1.3.0 vulnerable | "alice-thread" | Returns Alice |
| 2 | Injection attack | 1.3.0 vulnerable | { "$gt": "" } | Returns Bob |
| 3 | Same attack, patched | 1.3.1 patched | { "$gt": "" } | Rejected before any lookup |
| 4 | Normal request, patched | 1.3.1 patched | "alice-thread" | Returns Alice (fix is not over-broad) |
| 5 | Bob's real ID, patched | 1.3.1 patched | "bob-thread" | Returns Bob (the patch checks types, not ownership) |
Trial 5 is not a LangGraph bug. It shows what a type check cannot do: the ID is valid text, so deciding whether Alice may read it is the application's job.
node --version)setup.sh: macOS, Linux, or Windows with Git Bash or WSL.
The npm run commands also work from PowerShell.Install dependencies once with npm install.
The two saver versions are installed side by side as npm aliases (see
package.json):
"mongodb-saver-vulnerable": "npm:@langchain/[email protected]",
"mongodb-saver-patched": "npm:@langchain/[email protected]"
./setup.sh # start: check prereqs, start MongoDB, seed the data, serve the UI
./setup.sh test # run the five trials as integration tests
./setup.sh logs # show MongoDB's own log of each trial's query (add -f to follow)
./setup.sh stop # stop and remove ONLY this demo's MongoDB container + volume
Then open: http://127.0.0.1:4400. You should see "Local server ready" at the top right. Press Ctrl+C to stop the web server.
Without bash, the same steps are npm run db:up, npm start, npm test, and
npm run db:down. These use this project's Compose file (compose.yaml, project
name cs440-nosql-demo) and touch only this demo's container and volume.
Alice is the simulated requester. The chat message is always the literal text
Continue my chat. Only the identifier fields change between trials. Click the
numbered buttons in order, 1 to 5.
A trial finishes in milliseconds, so the page reveals the finished result one panel at a time, about a second apart. That pause is display pacing only, not query time.
Reset re-seeds the fictional data (Alice first, Bob last) and clears the
results table. Download evidence saves the last run as JSON. Every run is also
appended to evidence/runs.jsonl.
{ "$gt": "" } means "any non-empty thread_id", which broadens Alice's lookup
to every stored thread. The saver then applies .sort("checkpoint_id", -1) and
.limit(1), both visible in the captured query, so it returns the single newest
checkpoint. Bob is seeded last, so that checkpoint is Bob's.
Important nuance: a broadened lookup returns the latest matching checkpoint, not every record and not an arbitrarily chosen victim.
thread_id and checkpoint_nsLangGraph can save the state of a conversation so it can be restored
later. Each saved snapshot is a checkpoint. To find the right checkpoint, the
saver looks it up by a few identifier fields you pass in config.configurable:
thread_id: which conversation. Normally a plain string such as
"alice-thread".checkpoint_ns: checkpoint namespace. A sub-label used to separate
checkpoints within one conversation. For a simple top-level chat it is "".checkpoint_id (optional): which exact snapshot. If you don't provide
one, the saver returns the latest checkpoint for that thread.Version 1.3.0 dropped these values straight into a MongoDB filter. MongoDB treats
a field whose value is an object like { "$gt": "" } as an operator, not
as a literal id. Version 1.3.1 rejects any identifier field that is not a string.
graph.invoke(...) on a compiled
StateGraph, and each trial calls graph.invoke(...) again: the real
restoration path that internally calls the saver's getTuple.commandStarted monitoring event.Disclosed simplifications:
put / putWrites)
are suppressed while reads (getTuple / list) run the real, unmodified
code. This keeps the seeded data identical between runs. Seeding uses the fully
unwrapped saver, so the stored checkpoints are written by the genuine saver. See
src/saver.js (readOnly) and src/demo.js.src/fixture.js),
which go straight into the real graph.invoke. The demo does not expose a
free-form public endpoint. An app that forwards client-supplied IDs without
checking them is the affected scenario the advisory describes.find commands, so they appear in neither the captured
query nor the rows of ./setup.sh logs.The query panel shows the command the app captured. For evidence that does
not rely on the app's own word, compose.yaml starts mongod with
--slowms 0, so it logs every operation, and this prints those log entries:
./setup.sh logs # each trial's checkpoint query
./setup.sh logs -f # follow live
Lookups from seeding return nothing and are counted but not listed.