Skip to content
TOPIC

AI Security

Vulnerabilities, leaks, and security risks in AI products: agents reading untrusted content, session isolation, non-standard tool calling, and how to protect yourself.

AI securityAI vulnerabilityAI leakAI security riskprompt injectionagent isolation
[ FREQUENTLY ASKED QUESTIONS ]

FAQ

Q.01 What are the main security risks in AI agents?

The GitLost case showed GitHub agents leaking private repositories when processing malicious content. Claude Code had a possible session leak between accounts. Models can invent fields when calling tools. The main risk is treating agent output as trusted without security validation.

Q.02 How to protect AI agents from malicious content?

Isolate the agent execution environment. Validate all output before executing actions. Do not allow the agent to access resources beyond the necessary scope. Implement sandboxing for file and network operations. Monitor runtime with auditable evidence (like Halo). And never blindly trust model output.

[ RELATED CONCEPTS ]

Glossary

VIEW ALL →