Is smolagents safe?
- Python shell/command execution
- Unsafe deserialization
- Python dynamic code execution
What to do: Nothing here argues against installing it. Grant the capabilities it lists only if you expect the tool to need them.
smolagents is a PyPI package analyzed by SkillTotal's deterministic static scanner. The scan found no malicious indicators, though 1 risky construct is reported for review. It can: dynamic code execution, filesystem read, filesystem write, mcp tools detected, network egress and shell execution — capabilities are what the code can do, not a verdict on intent. Risk score 20/100 (low).
smolagents 1.26.0
Automated static-analysis result. It can contain false positives and false negatives, and is not a claim about the intent of smolagents's authors. Report a false positive.
Behavioral traits
How this component maps to the CSA agentic threat model. Descriptive — it never affects the risk score.
Findings (7)
It loads data with a format that can rebuild arbitrary objects (e.g. pickle, or unsafe YAML).
return pickle.loads(base64.b64decode(encoded_value[7:]))
return pickle.loads(base64.b64decode(data[7:]))
return pickle.loads(base64.b64decode(data))
Why it matters: Feeding such a loader untrusted data can execute code hidden inside that data.
Fix: Deserialize untrusted data with a safe format/loader: JSON, or yaml.safe_load / Loader=SafeLoader. Reserve pickle/marshal for data you fully control.
The code turns strings into live code at runtime (eval / new Function / exec).
exec(tool_code, module.__dict__)
Why it matters: If those strings aren't fixed and trusted, they become a way to run arbitrary code.
Fix: Avoid evaluating dynamically constructed code; if unavoidable, ensure the input is a trusted constant and never derived from external data.
The component can run operating-system commands or spawn processes.
subprocess.run([deno_path, "--version"], capture_output=True, check=True)
self.server_process = subprocess.Popen(
cmd,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
text=True,
)Why it matters: Powerful and often legitimate — confirm the commands aren't built from untrusted input.
Fix: Confirm the command and its arguments are fully controlled and not derived from untrusted input; avoid shell=True.
The component reads files from disk.
agent_dict = json.loads((folder / "agent.json").read_text())
tool_code = (folder / "tools" / f"{tool_name}.py").read_text()importlib.resources.files("smolagents.prompts").joinpath("toolcalling_agent.yaml").read_text()importlib.resources.files("smolagents.prompts").joinpath("structured_code_agent.yaml").read_text()importlib.resources.files("smolagents.prompts").joinpath("code_agent.yaml").read_text()tool_code = Path(tool_file).read_text()
Why it matters: Usually legitimate, but worth confirming it can't be steered into reading sensitive files.
Fix: Confirm which files are read and that paths cannot be influenced by untrusted input to reach sensitive locations.
The component writes or deletes files on disk.
with open(os.path.join(output_dir, "prompts.yaml"), "w", encoding="utf-8") as f:
with open(os.path.join(output_dir, "agent.json"), "w", encoding="utf-8") as f:
with open(os.path.join(output_dir, "requirements.txt"), "w", encoding="utf-8") as f:
with open(os.path.join(output_dir, "app.py"), "w", encoding="utf-8") as f:
shutil.copy(file_path, dest_path)
with open(self.runner_path, "w") as f:
shutil.rmtree(self.runner_dir)
file_path.write_text(content, encoding="utf-8")
with open(os.path.join(folder, "__init__.py"), "w"):
Why it matters: Usually legitimate, but worth confirming the paths can't be controlled by untrusted input.
Fix: Confirm which files are written/deleted and that paths cannot be influenced by untrusted input.
The component makes outbound network requests.
import requests
response = requests.get(self._path)
import requests
response = requests.get(base_url, params=params)
import requests
response = requests.get(self.endpoint, headers=self.headers, params=params)
import requests
response = requests.get(
"https://lite.duckduckgo.com/lite/",
params={"q": query},
headers={"User-Agent": "Mozilla/5.0"},
)import requests
response = requests.get(
"https://www.bing.com/search",
params={"q": query, "format": "rss"},
)import requests
response = requests.post(
"https://api.exa.ai/search",
headers={
"x-api-key": api_key,
"Content-Type": "application/json",
"x-exa-integration": "smolagents", …import requests
from requests.exceptions import RequestException
response = requests.get(url, timeout=20)
import requests
response = requests.get(self.server_url)
response = requests.post(self.server_url, json=payload, timeout=self.timeout)
import requests
from requests.exceptions import RequestException
r = requests.post(crate_kernel_endpoint, headers=headers)
if requests.get(f"{self.base_url}/api/kernelspecs?token={token}", timeout=2).status_code == 200:resp = requests.get(f"https://{host}/api/kernelspecs?token={token}")Why it matters: Usually legitimate, but confirm the destinations are expected and no sensitive data leaves.
Fix: Confirm the destination hosts are expected and that no sensitive data is sent off-host.
An MCP tool surface (manifest or tool definitions) was found.
source_code = get_source(self.forward).replace("@tool", "")# If tool was created using '@tool' decorator, it has only a forward pass, so it's simpler to just get its code
forward_source_code = forward_source_code.replace("@tool", "").strip()else: # If the tool was not created by the @tool decorator, it was made by subclassing Tool
f"Multiple @tool decorators found on function '{func_node.name}'. Only one @tool decorator is allowed."f"Multiple @tool decorators found on function '{func_node.name}'. Only one @tool decorator is allowed."f"Function '{func_node.name}' has decorators other than @tool. "Why it matters: Just context — review which tools it offers and their permissions.
Fix: Review the declared MCP tools and their permissions.
How attackers abuse these capabilities
Interactive labs on the attack class behind the rules above. They show the technique, not anything found in smolagents.
Check your own component
Run the same evidence-backed scan on any MCP server, agent skill, or package.
Scan your own componentHow we determine this: deterministic static analysis (regex + AST), evidence-anchored, no code execution. Methodology →