Threat Model
What PraisonAI trusts, what it does not, and which guard rails are not security boundaries
PraisonAI trusts you, the operator, and treats everything the model reads or writes as untrusted. This page shows where that line sits, so you know which settings protect you and which only prevent accidents.
Trust boundaries
The operator, their user-level config and environment variables, and code they install or register: packages, plugins, tools, hooks.
UntrustedModel output, tool results, fetched web content, project files in a repository you did not write, and anyone who connects to a PraisonAI server.
A vulnerability is untrusted input crossing to the trusted side without your consent: running code, reading or writing files you did not allow, reaching secrets, or acting through a server without authorisation.
Not a security boundary
| Feature | Why it is not a boundary |
|---|---|
Python code execution: ExecutionConfig(code_mode=...) set to "safe", "unsafe" or "isolated" | Runs on your machine as your OS user, in the agent process or a local subprocess. Restricted builtins, import filters and resource limits are guard rails. |
TypeScript codeMode tool | Runs in a Node.js vm context, which is not a sandbox. |
| String-based shell command filters | Allow or deny patterns over a command line can be worded around. |
A way around these is a hardening bug. Report it as a normal public issue, not a security advisory.
Run untrusted code in an isolating sandbox
When the model writes code, or the code comes from someone else, run it in a backend from praisonai-sandbox that isolates it from your machine.
pip install "praisonai-sandbox[docker]"| Backend | Where code runs |
|---|---|
docker | A local container |
sandlock | A Landlock and seccomp sandbox (Linux) |
e2b, modal, daytona, novita | A remote cloud sandbox |
ssh | A remote machine you control |
subprocess | Your machine, as your user: not isolation |
Network servers
Agents can be served over HTTP, MCP, A2A and AG-UI. Bind addresses differ between entry points, and some default to all interfaces, so set the host yourself.
- Bind to loopback
Use
host="127.0.0.1"unless other machines must reach the server. - Put a token in front of anything exposed
Use the server’s token option, or an authenticating reverse proxy.
from praisonaiagents import Agent
agent = Agent(instructions="Answer questions")
agent.launch(port=8000) # HTTP, binds to 127.0.0.1 by defaultAgent.launch() over HTTP binds to 127.0.0.1 by default. On any other address it requires a bearer token: set PRAISONAI_LAUNCH_AUTH_TOKEN, or one is generated and printed at start-up.
Reporting a vulnerability
What a report needs
- A working proof of concept against the latest release, in its default configuration.
- One report per root cause.
- A way around an earlier fix goes on the original advisory, not a new report.
- Scanner or AI-tool output without a tested proof of concept is closed.
Response targets
| Step | Target |
|---|---|
| Acknowledge | 3 working days |
| Decision | 14 days |
| Fix for a confirmed high or critical issue | 30 days |
Maintainers assign severity with CVSS 3.1, scored against the default configuration.
Other repositories
praisonai-platform and PraisonAI UI (aiui) take reports in their own repositories: praisonai-platform and PraisonAIUI.