Skip to main content
PromptGuard goes beyond access control to provide four governance capabilities that Chris Hood identified as missing from every AI security platform: agent identity, behavioral drift detection, tamper-evident audit trails, and runtime behavioral evaluation.

​
Why Governance Is Not Access Control

Access control answers “is this request allowed?” Governance answers harder questions:
  • Who is this agent? (Agent Identity)
  • Is it behaving the same way it did last week? (Behavioral Drift Detection)
  • Has anyone tampered with the audit trail? (Tamper-Evident Audit)
  • Can we prove compliance to an auditor? (Runtime Behavioral Evaluation)
PromptGuard handles all four.

​
Agent Identity

​
The Problem

Most AI security platforms accept a free-form agent_id string. Anyone with the project API key can claim to be any agent. There is no verification.

​
The Solution

PromptGuard offers verified agent credentials. Each agent is registered and receives a cryptographic secret (pgag_...) that authenticates requests. 
# Register an agent
curl -X POST https://api.promptguard.co/api/v1/agent/register \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_name": "customer-support-bot"}'
Response (secret shown only once): 
{
  "agent_id": "agent_a1b2c3d4e5f67890",
  "agent_name": "customer-support-bot",
  "agent_secret": "pgag_Kx9mP2qR...",
  "credential_prefix": "pgag_Kx9mP2"
}
Use the credential in subsequent requests: 
curl -X POST https://api.promptguard.co/api/v1/guard \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "X-Agent-Credential: pgag_Kx9mP2qR..." \
  -H "Content-Type: application/json" \
  -d '{"messages": [...]}'

​
Backward Compatibility

Requests without X-Agent-Credential still work. They are treated as unverified — the agent_verified field in responses and audit logs will be false.

​
Credential Rotation

curl -X POST https://api.promptguard.co/api/v1/agent/{agent_id}/rotate-credential \
  -H "Authorization: Bearer YOUR_API_KEY"
The old credential is immediately revoked and a new one is issued.

​
Behavioral Drift Detection

​
The Problem

An agent that normally uses search and calculator suddenly starts calling send_email and http_post exclusively. Per-request validation sees each call as individually safe. Only stateful behavioral analysis detects the drift.

​
The Solution

After an agent accumulates enough observations (configurable, default 50 tool calls), PromptGuard freezes a behavioral baseline — a snapshot of the agent’s normal tool-usage distribution. Every subsequent request is compared against this baseline using Jensen-Shannon divergence, a symmetric, bounded (0–1) measure with no division-by-zero issues. If the divergence exceeds the configured threshold (default 0.3), a BEHAVIORAL_DRIFT alert fires with:
  • Divergence score
  • Top changed tools (baseline vs. current distribution)
  • Recommended action

​
Example Alert

{
  "alert_type": "behavioral_drift",
  "severity": "warning",
  "message": "Behavioral drift detected (JS divergence=0.47)",
  "details": {
    "divergence": 0.47,
    "threshold": 0.3,
    "baseline_sample_size": 120,
    "changed_tools": {
      "send_email": {"baseline": 0.0, "current": 0.35, "delta": 0.35},
      "search": {"baseline": 0.6, "current": 0.15, "delta": -0.45}
    }
  }
}

​
Tamper-Evident Audit Trail

​
The Problem

An audit log that can be silently modified is not an audit log. Most platforms hash individual events but do not chain them — meaning a deleted or altered event goes undetected.

​
The Solution

Every audit event’s SHA-256 integrity hash incorporates the previous event’s hash (previous_hash), forming a cryptographic append-only chain. If any event is modified, inserted, or deleted, the chain breaks. 
Event 1: integrity_hash = SHA256(fields + previous_hash=null)
Event 2: integrity_hash = SHA256(fields + previous_hash=Event1.integrity_hash)
Event 3: integrity_hash = SHA256(fields + previous_hash=Event2.integrity_hash)

​
Verify the Chain

curl -X POST https://api.promptguard.co/dashboard/audit-log/verify-chain \
  -H "Cookie: session=YOUR_SESSION_COOKIE" \
  -d "start_date=2026-01-01T00:00:00Z" \
  -d "end_date=2026-04-10T00:00:00Z"

{
  "valid": true,
  "verified_count": 12847,
  "first_break_at": null
}

​
Runtime Behavioral Evaluation

​
The Problem

Compliance auditors need narratives, not data dumps. They need to understand what happened, what controls were in place, and whether those controls worked.

​
The Solution

PromptGuard generates governance reports with narrative sections for auditors: 
curl -X POST https://api.promptguard.co/dashboard/compliance/governance-report \
  -H "Cookie: session=YOUR_SESSION_COOKIE" \
  -d "framework=soc2&days=30"
The report includes:
SectionContents
Agent IdentityRegistered agents, verification rates, unverified access attempts
Behavioral DriftDrift alerts triggered, agents that drifted, resolution actions
Audit IntegrityChain verification result for the period
Security DecisionsBlock/allow/redact counts with explanation summaries
Incident TimelineChronological narrative of critical events
Supports SOC 2, EU AI Act, and NIST AI RMF frameworks.

​
Next Steps

Audit Logs

View and export audit events

Compliance

Security certifications and data handling

Agent Security

API reference for agent endpoints

Enterprise

Enterprise deployment guide