AI agents are no longer just answering questions. They’re writing code, managing infrastructure, executing financial transactions, and making decisions that affect real people. The governance gap isn’t theoretical anymore — it’s operational.

TealTiger v1.3.0 closes that gap with cryptographically verifiable governance for autonomous AI agents. Every governance decision now produces a tamper-evident receipt. Every agent identity is managed as a first-class principal. Every policy is deterministic, traceable, and auditable.

This isn’t incremental. v1.3 adds 7 new governance modules, 5 new providers, 3 platform adapters, and a complete evidence layer — while maintaining full backward compatibility with v1.2.

The Core Problem v1.3 Solves

Most AI governance today falls into one of two traps:

  1. “Trust but verify” (too late) — Log everything, review later. By the time you find the problem, the damage is done.
  2. “Block everything suspicious” (too noisy) — Use an LLM to judge another LLM. Non-deterministic, slow, and expensive.

TealTiger takes a third path: deterministic enforcement with cryptographic proof. No LLM in the governance path. Same input + same policy = same decision, every time. And every decision produces a verifiable evidence trail.

What’s New in v1.3

Non-Human Identity (NHI) Governance

Agents are principals, not ambient credentials. v1.3 introduces full identity lifecycle management:

  • Agent states: active, suspended, revoked — with immediate enforcement on status change
  • Scope enforcement: agents can only access resources within their declared scope
  • Zero Standing Privilege (ZSP): no permanent access. Agents get just-in-time grants that expire
  • Attestation: verify agent integrity before granting access
const engine = new TealEngineV13({
  nhi: {
    inventory: [
      {
        id: 'agent-billing-001',
        status: 'active',
        scopes: ['billing.read', 'billing.write'],
        environment: 'production',
        zsp: { enabled: true, max_grant_ttl: '5m' }
      }
    ]
  }
});

When an agent’s status changes to revoked, all subsequent requests are immediately denied with NHI_REVOKED — no propagation delay, no cache invalidation needed.

TealProof — Cryptographic Governance Receipts

Every governance decision now produces a cryptographic receipt:

  • Decision hash: SHA-256(decision + context + timestamp + policy_version + prev_hash)
  • Merkle tree: All decisions form a hash chain. Tampering with any decision invalidates the chain.
  • RFC 3161 timestamping: Merkle roots are anchored to an external TSA on a configurable schedule
  • Governance Passport: Per-agent summary of policy-compliant execution history

This means you can prove — cryptographically — that a specific governance decision was made, with a specific policy, at a specific time. Third-party auditors can verify this independently using the standalone Verification SDK (no TealTiger dependency required).

TealFlow — Declarative Governance Workflows

Governance policies often need orchestration: “if this happens, then do that, but only if approved by someone with this role.” TealFlow makes this declarative:

name: high-risk-action-approval
on:
  agent_action:
    risk_score_above: 80

jobs:
  require-approval:
    steps:
      - name: Notify security team
        uses: tealtiger/notify-webhook
        with:
          url: $
      - name: Wait for approval
        uses: tealtiger/require-approval
        with:
          approvers: ['security-team']
          timeout: '30m'

Org-level workflows set a floor that team-level workflows cannot weaken. If a team tries to remove a control that the org requires, TealFlow rejects it with WORKFLOW_FLOOR_VIOLATION.

TealClassifier — Local ML Without the Latency

v1.3 adds optional ML-based detection alongside the existing regex patterns:

  • ONNX Runtime: Models run locally, no API calls, no data leaves the process
  • ≤20ms inference: CPU-only, ≤512 tokens
  • Ensemble modes: regex_only (v1.2 behavior), ml_only, ensemble_union (block if either detects), ensemble_intersection (block only if both detect)
  • Graceful fallback: If ONNX fails to load, reverts to regex_only with a CLASSIFIER_FALLBACK event

Behavioral Drift Detection (TealDrift)

Agents change behavior over time — model updates, prompt drift, tool availability changes. TealDrift establishes statistical baselines and alerts when behavior diverges:

  • Tracks: refusal rate, response length, topic distribution per agent/provider/model
  • Configurable threshold (sigma-based)
  • Requires minimum samples before alerting (prevents false alarms on cold start)

Cost Governance That Code Can’t Override (TealMonitor v2)

Application code can set budgets. But what if a developer raises the limit? TealMonitor v2 introduces governance-owned ceilings:

  • Governance limits take precedence over application limits
  • Application code cannot raise limits above the governance ceiling
  • Anomaly detection: alerts when a request costs more than a configurable multiple of the rolling baseline
  • Reasoning-token budgets: separate tracking for chain-of-thought tokens

OWASP Agentic Top 10 — Zero Config

const engine = new TealEngineV13({
  policy_packs: ['owasp-agentic-top10']
});
// That's it. All 10 ASI risks covered.

Each control maps to a specific OWASP category identifier, produces structured evidence, and can be extended without modifying shipped files.

Platform Adapters

v1.3 ships native adapters for three enterprise agent platforms:

Platform Integration Style
AWS Bedrock Agents Lambda-backed guardrail adapter
AWS AgentCore Pre/post action governance plugin
Azure AI Agent Service Tool-call pipeline middleware

All adapters use the same TealEngineV13.evaluate() internally — identical decisions regardless of platform.

12 LLM Providers

Five new providers join the existing seven: DeepSeek, Groq, Together AI, HuggingFace TGI, and xAI (Grok). Total market coverage: 95%+.

Architecture: How It Fits Together

Request → Pre-Evaluation Stage → Module Evaluation → Post-Evaluation Stage → Response
              │                        │                       │
              ├─ FREEZE check          ├─ TealGuard v2         ├─ TealProof (hash)
              ├─ PLAN_ONLY check       ├─ TealSecrets          ├─ TealAudit (TEEC)
              ├─ NHI validation        ├─ TealMonitor v2       ├─ Response hooks
              ├─ Attestation           ├─ TealDrift            ├─ OTel spans
              ├─ ZSP grant check       ├─ TealClassifier       ├─ TealFlow triggers
              └─ Scope/env check       └─ TealRegistry v2      └─ SIEM export

The pre-evaluation stage short-circuits on deny — if a FREEZE rule matches, nothing else runs. This guarantees that emergency controls always take precedence.

Backward Compatibility

v1.3 is fully backward compatible with v1.2:

  • All v1.2 policy configurations work without modification
  • Identical Decision outputs for v1.2 configs
  • No v1.3 features configured = v1.2 behavior
  • All new config fields are optional with v1.2-equivalent defaults
  • TealClassifier defaults to regex_only
  • TealFlow is opt-in

Install

npm install [email protected]
pip install tealtiger==1.3.0
docker pull tealtigeradmin/tealtiger-typescript:1.3.0

What’s Next: v1.4 — Zero-Config Adoption

v1.3 is the most comprehensive governance release we’ve built. v1.4 will focus on making it effortless to adopt:

  • observe() mode — 1-line integration, instant visibility, no policies needed
  • Progressive disclosure: observe → suggest → enforce
  • Auto-baseline behavioral detection
  • Framework adapters for LangChain, CrewAI, AutoGen, LlamaIndex

TealTiger is open source (Apache 2.0). Star us on GitHub if you believe AI agents need governance, not just guardrails.