How to stop prompt injections and AI hallucinations at the OS level (CBS v1.6.2 release)

Description and use

<p><strong>Disclosure: I am the author of Cyber-Butler-Security (CBS).</strong></p> <p>If you run AI Agents (like DeepSeek Harness, AstrBot, or LangChain) locally on your machine, you've probably experienced the "Oh no" moment. </p> <p>Maybe the AI hallucinated and deleted a folder. Maybe a malicious prompt injected through a webpage told the agent to read your <code>~/.ssh</code> keys. </p> <p>Most solutions today try to fix this at the <strong>application layer</strong>: writing better prompts, hooking into the agent's API, or using a secondary LLM to judge the first LLM. </p> <p>But here is

Problem

<p><strong>Disclosure: I am the author of Cyber-Butler-Security (CBS).</strong></p> <p>If you run AI Agents (like DeepSeek Harness, AstrBot, or LangChain) locally on your machine, you've probably experienced the "Oh no" moment. </p> <p>Maybe the AI hallucinated and deleted a folder. Maybe a malicious prompt injected through a webpage told the agent to read your <code>~/.ssh</code> keys. </p> <p>Most solutions today try to fix this at the <strong>application layer</strong>: writing better prompts, hooking into the agent's API, or using a secondary LLM to judge the first LLM. </p> <p>But here is

Outcome

<p><strong>Disclosure: I am the author of Cyber-Butler-Security (CBS).</strong></p> <p>If you run AI Agents (like DeepSeek Harness, AstrBot, or LangChain) locally on your machine, you've probably experienced the "Oh no" moment. </p> <p>Maybe the AI hallucinated and deleted a folder. Maybe a malicious prompt injected through a webpage told the agent to read your <code>~/.ssh</code> keys. </p> <p>Most solutions today try to fix this at the <strong>application layer</strong>: writing better prompts, hooking into the agent's API, or using a secondary LLM to judge the first LLM. </p> <p>But here is

Evidence

How to stop prompt injections and AI hallucinations at the OS level (CBS v1.6.2 release)

Source