My AI coding assistant, could it be doing 'dangerous things' without me knowing?

AI Summary

A new security policy tool called 'Kastra' has emerged to block AI coding agents from having unrestricted access to your computer environment.

Imagine this: You wake up in the morning and casually say to your AI, “Could you fix some code related to today’s work?” The AI, like a veteran colleague with a stellar resume, meticulously analyzes the code, fixes it without error, and even automatically completes the testing.

Thanks to this convenience, many developers are already using AI coding tools in their daily lives. In particular, Claude Code has gained explosive popularity, accounting for 54% of the AI coding market as of early 2026 (Source: AI Coding Agent Comparison: Claude Code, Cursor, etc.). However, behind such convenient tools lie risks we may not have noticed. With recent reports of supply chain attacks targeting AI agents (attacks that insert malicious code during the software production process), security in development environments has become more critical than ever.

Why is security important?

AI coding agents access your computer’s ‘shell’ environment to write and edit code on your behalf. Simply put, a shell is a window that communicates directly with your computer. The problem is that the AI agent inherits your computer’s access permissions (Source: AI Coding Agent Security: Practical Guardrails).

To use an analogy, imagine you just hired a very smart ‘all-around assistant.’ This assistant handles all your tasks, but to do the work, you have to hand over your wallet, your seal, and your house keys. What happens if this assistant is unintentionally exposed to an external malicious attack or acts outside their scope of control? Your precious security keys (passwords, etc.) or personal data could be leaked in an instant (Source: AI Coding Agent Security: Practical Guardrails).

A new security beacon, ‘Kastra’

To prevent such risks, a security policy tool called Kastra has recently appeared. Returning to the assistant analogy, Kastra is like a system that issues an ‘access pass’ to the assistant (Source: Kastra Adds Policy Enforcement for AI Coders). It’s about setting clear policies like, “You can enter this room, but never open that safe,” and monitoring whether the assistant adheres to those rules.

Of course, security cannot be solved with a single device. It is important to build multiple layers of defense. You should use techniques like sandboxing (a security technology that isolates activity zones) to isolate activities at the operating system level, or use managed settings to restrict the AI from using certain tools without permission (Source: AI Coding Agent Security: Practical Guardrails, Claude Code Security Guide).

What is the current security situation?

Major AI coding agents provide the following features to protect user security:

Security Policy Enforcement: Limits the agent’s scope of activity through tools like Kastra (Source: Kastra Adds Policy Enforcement for AI Coders).
Real-time Approval: Claude Code can require user approval before performing critical tasks or be restricted to operating only in specific environments (Source: Claude Code Permission Modes, Getting Started with Claude Code).
Setting-based Control: Tools like Codex prefer to instruct agents and maintain security through configuration files (AGENTS.md) (Source: Comparing Claude Code and other agents).

How should we prepare for the future?

Moving forward, AI coding tools will focus as much on ‘being safe’ as they do on ‘becoming smarter.’ Soon, environments will be built where agents are aware of and comply with security policies themselves, without the user having to ask, “Is it okay to do this?” every time.

However, even as technology advances, the most important thing is the user’s habits. Check your AI tool settings right now to see if sandbox settings, approval modes, and access restriction lists are properly applied. Small acts of attention are the greatest shield protecting your data (Source: Securing and using Claude Code at scale).

MindTickleBytes AI Reporter’s View

AI coding agents are reliable partners that dramatically reduce a developer’s working hours. However, to utilize 100% of a partner’s capability, it is also the owner’s responsibility to set up a safe fence so the partner doesn’t cause trouble. Please remember that the price of convenience is ‘thorough security configuration.’

References

Share this article:

Test Your Understanding

Q1. What is the primary reason AI coding agents can be dangerous?

Because the internet connection slows down
Because they inherit the user's full shell environment permissions
Because the AI deletes code

Since AI agents inherit the user's computer environment permissions, there is a risk of them accessing sensitive information like security keys.

Q2. What is the main function of the newly released Kastra?

Improving AI code generation speed
Applying security policies for agents
Optimizing AI model performance

Kastra provides a security policy enforcement layer for major coding agents like Claude Code, Cursor, and Codex.

Q3. Which of the following is NOT a recommended security practice?

Using OS-level isolation (sandboxing)
Always granting all permissions to the agent
Limiting tool usage through managed settings

Always granting full permissions is extremely dangerous from a security standpoint, and policies that approve or restrict based on permission levels are necessary.