The Security Guide

Attack Surface

AI agent harnesses introduce a unique attack surface. Key vectors:

Prompt Injection

Malicious instructions embedded in code, comments, issues, or documentation that attempt to manipulate agent behavior.

Defenses:

ECC’s InsAIts security hook (opt-in via ECC_ENABLE_INSAITS=1)
Input validation in hooks
Secret scrubbing in hook outputs

Tool Misuse

Agents have access to powerful tools (file write, bash execution, git operations). Uncontrolled tool use can lead to:

Accidental file deletion
Credential exposure
Unauthorized git pushes

Defenses:

PreToolUse hooks that validate commands
Git push confirmation hooks
Dev server blockers (prevents npm run dev outside tmux)

Context Poisoning

Malicious content in the context window that influences agent decisions.

Defenses:

Session isolation
Context compaction with key-extraction
Verification loops

Sandboxing

Run agents in isolated environments:

Docker containers - Full isolation for untrusted operations
Git worktrees - Filesystem isolation for parallel tasks
Tmux sessions - Process isolation for long-running commands

Security Rules

ECC installs security rules for each language ecosystem:

Input validation patterns
Authentication best practices
SQL injection prevention
XSS defense
Secret management
Dependency auditing

Risk Quantification

Use the harness audit command to score your security posture:

/harness-audit

This evaluates hook coverage, rule compliance, and MCP configuration security.

AgentShield

The ecc-agentshield npm package provides additional security layers for production agent deployments. It includes runtime monitoring, anomaly detection, and automated response capabilities.