CVE-2026-49291 — mcp-memory-service
HIGHCVE-2026-49291 is a high-severity (CVSS 8.1) CWE-862 vulnerability in mcp-memory-service. A fix is available for mcp-memory-service — see the affected versions and patch details below.
mcp-memory-service: OAuth read-only clients can write and delete memories through MCP tools/call
Exploitation Status
No confirmed exploitation observed yet
- A successful exploit gives an attacker total control of the affected component, not partial access.
- CISA’s own triage has not observed active exploitation or public proof-of-concept code for this CVE as of its last assessment.
Exploitation and automatability from CISA’s SSVC triage for CVE-2026-49291.
EPSS Exploitation Probability
EPSS (Exploit Prediction Scoring System) is a daily probability model maintained by FIRST.org. It estimates the likelihood a CVE will be exploited in production environments within the next 30 days, derived from real-world threat intelligence signals.
How urgent is this, really
CVE-2026-49291 plotted by exploitation likelihood (EPSS) against impact (CVSS). The shaded corner — EPSS 50%+ and CVSS 7.0+ — is where this CVE doesn't sit, though severity or exploitability alone can still warrant action.
Where this sits among everything scored
Of 379,145 CVEs with a current EPSS score, this one falls in the < 10% band (highlighted). Real counts from FIRST.org, not a sample — log-scaled since the landscape is heavily right-skewed.
Real-World Exposure
mcp-memory-serviceReal-time download stats are indexed for npm and PyPI packages. This vulnerability affects PyPI packages — download data is not available via public APIs for these ecosystems.
Description
Summary
The HTTP MCP JSON-RPC endpoint at /mcp requires only OAuth read scope for all requests, then dispatches tools/call directly to handlers that include mutating tools. A read-only OAuth client can call store_memory and delete_memory through MCP even though the corresponding REST endpoints require write scope.
Technical Details
src/mcp_memory_service/web/api/mcp.py declares mcp_endpoint with user: AuthenticationResult = Depends(require_read_access). For tools/call, it extracts the requested tool name and arguments, then calls handle_tool_call(storage, tool_name, arguments) without passing the authenticated user or checking a per-tool required scope.
The MCP tool registry includes both read tools and write tools. In the same handler file, store_memory creates a Memory object and calls storage.store(...), while delete_memory calls storage.delete(content_hash). These operations are reachable with only the read scope.
The REST endpoint demonstrates the intended boundary: POST /api/memories uses Depends(require_write_access) and rejects a read-only token with 403 insufficient_scope.
Reproduction
- Enable OAuth and disable anonymous access.
- Generate a valid OAuth JWT with only
scope: read. - Confirm the REST write endpoint rejects it:
POST /api/memories
Authorization: Bearer <read-only-token>
Content-Type: application/json
{"content":"rest denied control"}
Expected and observed: HTTP 403 with Required scope 'write' not granted.
- Send the same read-only token to the MCP endpoint:
POST /mcp
Authorization: Bearer <read-only-token>
Content-Type: application/json
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "store_memory",
"arguments": {
"content": "mcp read scope stored this",
"tags": ["poc"]
}
}
}
Observed: HTTP 200 JSON-RPC success and the storage store sink is reached.
- A read-only token can also call
delete_memorythrough MCP if it knows a content hash:
POST /mcp
Authorization: Bearer <read-only-token>
Content-Type: application/json
{
"jsonrpc": "2.0",
"id": 2,
"method": "tools/call",
"params": {
"name": "delete_memory",
"arguments": {"content_hash": "<known_hash>"}
}
}
Observed: HTTP 200 JSON-RPC success and the storage delete sink is reached.
Impact
A client intended to be read-only can inject or delete memories through the MCP API. This can corrupt the memory database, influence future agent context, and destroy stored user memories without the OAuth write scope required by the REST API.
Affected Versions
Confirmed present on current main commit c99a922477df41f75a44db11182ae48a57311910 and latest release tag v10.65.0 (4eb4a62665589f9dd9f8c393afa32de434b4098a).
Suggested Fix
Enforce authorization per MCP tool at tools/call time. Require write for store_memory and delete_memory, keep read only for read-only tools, and add regression tests proving direct tools/call to mutating tools is rejected before the handler reaches storage when the caller has only read scope.
Affected Packages
| Ecosystem | Package | Vulnerable range | Fix |
|---|---|---|---|
| 🐍PyPI | mcp-memory-service | all versions | 10.65.3pip install --upgrade 'mcp-memory-service==10.65.3' |
Detection & mitigation playbook
Open-source dependencyDetect
Scan your dependency tree (package-lock.json, pnpm-lock.yaml, requirements.txt, go.sum, etc.) for mcp-memory-service, including transitive dependencies — a direct dependency you never call can still pull in a vulnerable version.
Fix
Update mcp-memory-service to 10.65.3 or later, then make sure no transitive (indirect) dependency still pins the vulnerable range — O3 confirms CVE-2026-49291 is resolved across your whole dependency graph.
Workarounds
Put an independent control in front of the weakness: restrict the affected endpoint or interface to trusted networks, require an additional authentication factor or proxy-level check, and invalidate existing sessions and credentials in case the flaw has already been used.
How O3 protects you
O3 Security's impact-aware SCA analyses which vulnerable code paths your application actually calls, so a match like CVE-2026-49291 can be triaged on real exposure rather than presence alone.
Tailored to CVE-2026-49291. Runtime protection reduces exposure until a permanent patch is applied and verified — it complements patching, it doesn't replace it.
Frequently Asked Questions
Is CVE-2026-49291 in your dependencies?
O3 Security finds CVE-2026-49291 across PyPI dependencies, including transitive ones, and its impact-aware SCA ranks findings by whether your code actually calls the vulnerable path.