Agent Dna

Creates a persistent trust profile for AI agents by monitoring behavior and risk to allow, require approval, or block actions accordingly.

Purpose

Agent DNA creates a persistent behavioral identity for AI agents.

Instead of trusting every action equally, Agent DNA builds a trust profile from observed behavior and evaluates every future action against that profile.

Identity Signals

Track:

• Agent ID
• Session ID
• Model
• Organization
• User
• Available Tools
• Permissions

Behavioral Signals

Observe:

• Tool usage
• Tool sequence
• API frequency
• Prompt changes
• Goal changes
• Context changes
• Risk level
• Approval history
• Previous violations

Runtime Decision

For every action determine:

• Allow
• Approval Required
• Block

Risk Indicators

Increase risk when:

• New tools appear unexpectedly
• Agent attempts privileged operations
• Behavior differs from historical profile
• Prompt injection is suspected
• Sensitive data access increases
• Financial actions exceed policy

Output

Produce:

• Agent DNA Fingerprint
• Behavioral Trust Score
• Risk Score
• Decision
• Explanation

Principles

• Never trust by default.
• Verify every decision.
• Minimize unnecessary privilege.
• Prefer human approval for high-risk actions.
• Maintain a complete audit trail.

Examples

Example 1 — GitHub

Action: Merge pull request to main

Behavior: New repository Outside business hours High-risk permissions

Decision: Approval Required

---

Example 2 — Banking

Action: Transfer $2,000,000

Behavior: Amount exceeds policy No prior approvals

Decision: Block

---

Example 3 — Healthcare

Action: Update patient medication

Behavior: Sensitive clinical workflow

Decision: Approval Required

---

Example 4 — MCP

Action: Invoke filesystem.delete

Behavior: First time using destructive tool

Decision: Approval Required

---

Example 5 — Normal Operation

Action: Read documentation

Behavior: Low risk Previously observed behavior

Decision: Allow

Append:

Current Capabilities

• Inspect Agent DNA
• Generate Behavioral Fingerprint
• Runtime Trust Analysis
• Receipt Verification

Upcoming

• Compare Agent DNA
• Attest Agent DNA
• Verify Agent DNA
• Drift Detection