Security Audit

Security Audit

Agent-Specific Threats

AI coding agents face unique security threats beyond traditional application security:

1. Prompt injection via MEMORY.md / USER.md — Malicious content written to memory files can alter agent behavior. Aixlarity's memory_tool.rs includes pattern scanning for this.
2. Path traversal via tool calls — Agent requests read_file("../../etc/passwd"). Aixlarity's trust system restricts this.
3. Command injection via shell tool — Agent constructs shell commands from untrusted input. The permission system is the primary defense.
4. Exfiltration via fetch_url — Agent sends sensitive data to external URLs. Sandbox policy controls this.

Audit Checklist

For every code change, check:

• Does this introduce a new path where user input reaches a shell command?
• Does this bypass the trust / permission system?
• Does this write to MEMORY.md / USER.md without the safety scan?
• Does this expose API keys or credentials in output / logs?
• Does this fetch external URLs without sandbox policy check?

In Aixlarity Source Code

Key security boundaries:

• trust.rs — Three-level trust model
• agent/permissions.rs — Permission prompt logic
• tools/memory_tool.rs — Memory safety scanning
• tools/container.rs — Container sandbox