What is the biggest security problem in AI coding agents?

AI coding agents are getting access to repositories, terminals, databases, and deployment pipelines.

In your opinion, what's the biggest security risk?

• Secret leaks

• Dangerous tool calls

• Prompt injection

• Dependency attacks

• Excessive permissions

• Audit and compliance gaps

I'm building ShadowPaste to explore these problems, and I'm interested in hearing how other developers think about AI security.

What's the biggest challenge you've faced?