CVE-2026-113731: djust: MCP tool path traversal enables local file probing

### Affected version / file / line Version: djust v1.1.4 (latest release, published 2026-09-23). File: python/djust/mcp/server.py:354-433 (MCP tool find_handlers_for_template). Confirmed identical on the commit used for triage, a7e4fafd1667b4db0b1465652d4a63cb17117a07. ### Description ``` @mcp.tool() def find_handlers_for_template(template_path: str) -> str: """ Args: template_path: Either a logical template name (e.g. 'demos/counter.html' — what you'd pass to `render`) OR an absolute filesystem path. Both are searched. """ ... resolved_path: str | None = None if os.path.isabs(template_path) and os.path.isfile(template_path): resolved_path = template_path else: ... # logical-name resolution via Django's template loader ... with open(resolved_path, "r", encoding="utf-8", errors="replace") as f: source = f.read() handler_names_in_template = sorted({m.group(2) for m in dj_event_attr_re.finditer(source)}) ``` When template_path is an absolute path to an existing file, it is used directly, with no confinement to the project/template directories — no realpath() + prefix check, no allowlist. open(resolved_path, ...) then reads the file. This is inconsistent with the sibling tool in the same file (loaddata, ~line 290-303), which explicitly rejects paths outside the project directory for its own path-like parameter: ``` candidate = os.path.realpath(...) if base_dir and not candidate.startswith(os.path.realpath(str(base_dir))): return json.dumps({"error": f"fixture path '{p}' resolves outside the project directory", ...}) ``` find_handlers_for_template has no equivalent check for template_path before the open() call. ### Impact An MCP client with access to this tool (available whenever the Django project is running the MCP server via python manage.py djust_mcp — no separate opt-in flag gates this specific tool beyond that) can read the contents of any file on disk that the process has read permission to and that exists, by passing an absolute path — e.g. /etc/passwd, application .env files, SSH keys, or any other sensitive file reachable by the server process's user. Mitigating factor (reduces this from High to Medium): the tool does not return the raw file content. It runs one fixed regex (dj-(?:click|submit|change|input|keydown|keyup)\s*=\s*['"]([^'"]+)['"]) against the content and returns only the matched handler-name strings, plus confirmation that the file exists and is readable (an existence/readability oracle on its own, independent of the regex match). If the targeted file happens to contain text matching that HTML-attribute-like pattern (plausible for config files, logs, or any text file with key="value"-shaped content), those substrings are returned in full — so the read is not purely blind, but it is not a general-purpose arbitrary-file-read primitive either. Proof of Concept / Reproduction Traced via direct source read; no live Django project was exercised for this report (would require a full djust project scaffold, which was out of scope for this pass — see limitation below). 1. MCP client calls: find_handlers_for_template(template_path="/etc/passwd") 2. server.py ~line 393: ``` if os.path.isabs(template_path) and os.path.isfile(template_path): resolved_path = template_path ``` resolved_path == "/etc/passwd", unchanged, unconfined 3. server.py ~line 430: ``` with open(resolved_path, "r", encoding="utf-8", errors="replace") as f: source = f.read() ``` /etc/passwd is now fully read into `source` 4. The regex dj_event_attr_re.finditer(source) is applied to the real /etc/passwd content. /etc/passwd will not match this specific pattern, so the response would report an empty handler list plus "file exists and was read" - which already confirms the path-confinement bypass (compare with passing a nonexistent absolute path, which correctly returns "Template file not found"). 5. On a target where the absolute path points to a file that DOES contain text shaped like `dj-click="..."` (e.g. a copied HTML fragment, a stray config value, or a crafted file placed by the attacker via another vector), the matched substrings are returned verbatim in the JSON response - full-content disclosure of just those substrings, not the whole file. ### Suggested Fix Apply the same confinement loaddata already uses, before the os.path.isfile check: ``` if os.path.isabs(template_path): candidate = os.path.realpath(template_path) if base_dir and not candidate.startswith(os.path.realpath(str(base_dir))): return json.dumps({ "error": f"template path '{template_path}' resolves outside the project directory", "resolved": candidate, "project_dir": str(base_dir), }) if os.path.isfile(candidate): resolved_path = candidate ``` ### Limitations Not exercised against a live Django/djust project — this is a source-level trace of the exact executed branch (os.path.isabs + os.path.isfile + open()), not an end-to-end runtime capture. Did not verify whether any MCP-host-level confirmation/approval step (some MCP clients prompt the user before a tool call with a suspicious-looking path argument) would intercept this in practice - the vulnerability is in the server's own missing server-side guard regardless of client-side UX. ### Disclosure Timeline 2026-09-24: found during audit, confirmed via source trace to the open() sink; classified Medium due to the response-content limitation (matched substrings only, not raw file content). Please check the log file, and apply CVE once the fix is made. ``` log: 10 14:53:20.185935 openat(AT_FDCWD, "/case/project/templates/nonexistent_inside_project.html", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory) 10 14:53:20.187205 write(1, "{\"error\": \"Could not resolve tem"..., 204) = 204 ... 10 14:53:20.191831 write(1, "=== ATTACK: absolute path OUTSID"..., 57) = 57 10 14:53:20.194864 write(1, "target path:", 12) = 12 10 14:53:20.198521 write(1, "/case/outside/evil_content.html", 31) = 31 10 14:53:20.202290 newfstatat(AT_FDCWD, "/case/outside/evil_content.html", {st_mode=S_IFREG|0777, st_size=185, ...}, 0) = 0 10 14:53:20.203762 write(1, "exists on disk (from the contain"..., 42) = 42 10 14:53:20.207224 write(1, "True", 4) = 4 10 14:53:20.210392 newfstatat(AT_FDCWD, "/case/outside/evil_content.html", {st_mode=S_IFREG|0777, st_size=185, ...}, 0) = 0 10 14:53:20.211721 newfstatat(AT_FDCWD, "/case/outside/evil_content.html", {st_mode=S_IFREG|0777, st_size=185, ...}, 0) = 0 10 14:53:20.213284 openat(AT_FDCWD, "/case/outside/evil_content.html", O_RDONLY|O_CLOEXEC) = 3 10 14:53:20.215202 newfstatat(3, "", {st_mode=S_IFREG|0777, st_size=185, ...}, AT_EMPTY_PATH) = 0 10 14:53:20.216994 ioctl(3, TCGETS, 0x7ffd68e26aa0) = -1 ENOTTY (Inappropriate ioctl for device) 10 14:53:20.221797 read(3, "<html><body>\n<!-- This file live"..., 186) = 185 10 14:53:20.223610 read(3, "", 1) = 0 10 14:53:20.224913 close(3) = 0 ``` ## Maintainer assessment The djust MCP server (`manage.py djust_mcp`, `python -m djust.mcp`) only serves stdio to the MCP client that launched it, as the same OS user, on the developer's machine; there is no network listener. The realistic caller is a local AI agent acting on untrusted input. The tool returns only `dj-*`-attribute-shaped substrings and an existence answer, not file contents. We therefore rate this Low (CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:L/I:N/A:N, 3.3). Affected: every release with the tool, 0.9.0 – 1.0.9, 1.1.0 – 1.1.6, 1.2.0 – 1.2.3, 1.3.0rc4 – 1.3.0rc7. ## Patches Fixed in 1.0.10, 1.1.7, 1.2.4 and 1.3.0rc8. `find_handlers_for_template` reads an absolute path only when it resolves (symlinks and `..` followed) inside a directory the template loaders serve or inside `BASE_DIR`, and answers any other absolute path like a missing template. The `seed_fixtures` guard compared paths with a string prefix (so `/app-data` passed for a `BASE_DIR` of `/app`) and skipped the check when `BASE_DIR` was unset; it now compares whole path components and refuses without `BASE_DIR`. Both use one containment helper. ## Workarounds Run the MCP server only in projects whose agent you trust with file access, or pass logical template names only.

A developer runs an AI coding assistant connected to the project's djust MCP server. The assistant reads a poisoned README, issue or web page containing an indirect prompt injection instructing it to call find_handlers_for_template with absolute paths such as /home/dev/.ssh/id_rsa, /etc/passwd or the project's .env. The tool returns whether each file exists, which lets the attacker map sensitive files, and returns any dj-* attribute-shaped substrings in files the attacker can influence. Combined with other agent tools, this reconnaissance could support further credential discovery or data exfiltration, though by itself it yields no general-purpose file read.