Coding agents and prompt injection
A coding agent does what the model tells it. That makes the model's answer a security boundary.
Because of the risk of prompt injection, every coding agent runs in the Certified pool, always. A request that declares tools, carries tool calls or tool results, or is sent with a person's sign-in token, which only hi and HIPS Code use, runs only in the Certified pool, whatever client sends it (Requests that can act).
Prompt injection is text that makes a model act against your instructions. It reaches a model in two ways:
- 01
Through what the agent reads. A library, a web page, a document or a tool result can hold instructions written for the model. The agent reads them as part of its work.
- 02
Through the machine that runs the model. Whoever controls that machine could change the prompt before the model reads it, or change the answer before you get it.
For a chat, the result is a wrong answer. For a coding agent, the answer holds tool calls: commands and file edits that run on your computer, next to your source code and your credentials.
- Coding agents run only in the Certified pool. Every request that can act through tool calls is placed in the Certified pool and is never sent to a GPU in the Community pool. HIPS does not offer coding-agent models in the Community pool. In the Certified pool the machines that run the model are certified by HIPS's compliance team, so the second way above is closed.
- Guardrails detect injected instructions. A content guardrail checks what goes to the model, tool results included, for text that tries to redirect it, and can flag, remove or block it. That addresses the first way above. Read Guardrails.
- Keys are redacted on the way to a provider's GPU. Before a request reaches a GPU a provider owns, HIPS recognizes and redacts API keys, tokens, private keys and the values of environment variables in its text.
- Nothing is stored. HIPS does not store, log or train on your code, your prompts or the answers. Read What HIPS stores.
Install hi; it signs you in on its first run (Install hi, with Homebrew and Windows). From hi 0.1.7, every request of a session hi starts runs in the Certified pool, with fallback off, whatever the workspace's tool defaults say (Coding agents). Coding agents and tool calls are on the paid plans (Requests that can act).
Shell
curl -fsSL https://hips.com/hi.sh | shhi claudeYour own agent, or any app that sends tool calls: name the Certified pool in the request too, so the code says what it does.
import osfrom openai import OpenAI client = OpenAI( api_key=os.environ["HIPS_API_KEY"], base_url="https://ai.hips.com/v1", # A request with tools runs only in the Certified pool default_headers={"x-hips-pool": "certified"},)result = client.chat.completions.create( model=os.environ["HIPS_MODEL"], messages=[{"role": "user", "content": "List the files in this folder."}], tools=[{"type": "function", "function": {"name": "list_files", "parameters": {"type": "object", "properties": {}}}}],)print(result.choices[0].message.tool_calls)- Start your coding agent through hi, or use HIPS Code once it is released.
- Keep your agent's permission prompts on, and use its sandbox where it has one.
- Read a command before you approve it, above all one that sends data out, installs a package or changes your shell or git configuration.
- Keep credentials out of the folder the agent works in.
Related sheets
- Sheet
- 03 of 21
- Detail
- B
- Updated
- 2026-11-17