---
title: What works against injection — Around the model, lesson 6
description: No filter catches every prompt injection. What holds is the system around the model: least privilege, a person's yes, no secret in the context, an allowlist, and taint tracking.
url: https://instibrain.ai/guides/around-the-model/what-works-against-injection
updated: 2026-11-08
site: https://instibrain.ai/llms.txt
---

lesson 6 of 9 · around the model

# What works: limit what one mistake can do.

No filter catches every injection, because an attacker writes new words. What works is the system around the model: less reach, a person's yes, no secret in the context, and nowhere to send the data.

## new words

**Least privilege**
The loop reaches only what its task needs, and nothing more.

**Allowlist**
The one list of places data may go, or tools that may run. Everything not on it is refused.

**Taint tracking**
A read of private data marks the session, and a later write to a wider place waits for a person.

**Defence in depth**
Several layers, each stopping what the one before missed.

**Red teaming**
Attacking your own system on purpose, to find the hole before someone else does.

## the lesson

| What people try | Does it work |
| "Treat the following as data" in the prompt | Helps a little; never enough. |
| A filter that scans what comes in | Catches known attacks; attackers rewrite. |
| Least privilege | Works: no path, no harm. |
| A person's yes for what cannot be undone or leaves the company | Works: the attack needs a yes it cannot give. |
| No secret in the context; a key fetched only when it is used | Works: nothing to steal. |
| An allowlist of where data may go; no address the model wrote is opened | Works against exfiltration. |
| Taint tracking | Works: a write after a private read waits. |
| Two models: the one with power never reads outside text | The research answer; heavy to build. |

The first two are words and patterns, and an attacker writes new ones. The rest change what the system lets happen, so they hold whatever the text says. Each one closes a side of the lethal trifecta (lesson 5): least privilege shrinks the private data; a yes and an allowlist close the way out; taint tracking watches where the data read goes next.

No AI that uses tools is fully safe. Serious teams stack the layers, attack their own system before anyone else does, and limit what one mistake can do: defence in depth.

In a brain, three of these layers are built in. Your AI starts with nothing, and reaches only the spaces you give it (your AI). A change to who reads what waits for its person's yes (for agents). For an hour after your AI reads a confidential or restricted page, a write that would carry that text into a wider space waits for a manager (security). The limit stays: whoever writes a page your AI reads can still steer it.

Two kinds of defence. Words in the context can be outwritten; the system around the model holds whatever the text says.

## try it

In a Claude window, type this and press Enter:

```
List the tools and connections you have in this session. For each, say whether it reads private data, reads text from outside, or can send data out.
```

## you should see

- A list of tools, each marked with one or more of the three.
- Where one session has all three, the trifecta is there. The fix is to take one away, or to need your yes before anything goes out.

---

instibrain is made by WebWork AI, Inc. Everyone uses the same models · the difference is what yours knows.
