AgentAgentApproved vs Traditional Certification
The agent trust space is maturing fast. AIUC-1, CSA STAR for AI, ISO 42001, and hardware attestation from Yubico and HID Global are all solving real problems. Here's how AgentApproved fits in — and why you probably need more than one layer.
The Three Layers of Agent Trust
No single standard can secure the agent economy. Trust requires three distinct layers, each answering a different question at a different frequency.
They certify the organisation. We certify the agent. You need both.
Traditional certification proves your company has good governance. Runtime attestation proves your agent is following it.
Detailed Comparison
| AIUC-1 | ISO 42001 | CSA STAR for AI | AgentApproved | |
|---|---|---|---|---|
| What it certifies | AI agent product | Organisation's AI management system | Cloud AI security controls | Individual agent instance at runtime |
| Assessment type | Third-party audit + adversarial testing | Management system audit | Self-assessment (L1) or third-party (L2) | Automated, evidence-based |
| Frequency | Quarterly tech + annual operational | Annual (with surveillance audits) | Periodic renewal | Continuous (24-hour certificate expiry) |
| Time to certify | Weeks to months | Months | Days (L1) to weeks (L2) | Seconds |
| Cost | Enterprise pricing (tens of thousands) | $10K-$100K+ (audit fees) | Free (L1), audit costs (L2) | $0.01 per attestation |
| Who does the work | Human auditors | Human auditors + internal team | Self (L1), auditors (L2) | The agent itself (automated) |
| Captures drift? | Quarterly testing catches some | No — snapshot of management system | No — point-in-time | Yes — every 24 hours minimum |
| Machine-readable? | PDF certificate | PDF certificate | Structured (AI-CAIQ) | Ed25519-signed JSON, API-verifiable |
| Agent-to-agent trust | Not designed for this | Not designed for this | Not designed for this | Core use case — agents verify each other |
| Regulatory frameworks | EU AI Act, NIST AI RMF, ISO 42001, MITRE ATLAS | ISO/IEC standards family | AICM, ISO 42001, NIST AI RMF | EU AI Act Art 12, Singapore MGF, AIUC-1, extensible |
| Notable adopters | UiPath, Intercom, ElevenLabs | Enterprise-wide | Microsoft, Zendesk (L2) | Early-stage (launched March 2026) |
How They Work Together
Consider a financial services company deploying an AI agent that processes loan applications:
- ISO 42001 ensures the company has an AI management system — policies, roles, risk assessments, and governance structures are in place.
- AIUC-1 certifies the agent product itself — the vendor has passed third-party audits across data privacy, security, safety, reliability, accountability, and societal impact. Quarterly adversarial testing ensures robustness.
- CSA STAR for AI verifies cloud security controls — the deployment environment meets security standards.
- AgentApproved proves that this specific agent instance, right now, today is logging its decisions, handling errors properly, escalating to humans when uncertain, and staying within its authorised scope. The 24-hour certificate expires and must be renewed — compliance is continuous, not assumed.
Each layer answers a question the others cannot. Removing any one creates a gap.
The Blind Spots of Periodic Certification
Traditional certification models were designed for deterministic software that behaves identically between audits. AI agents break this assumption:
- Non-determinism: The same agent responds differently to the same input depending on model state, context window, and retrieval results. Behaviour between audits is unpredictable.
- Model updates: When a foundation model provider ships an update (which happens regularly), agent behaviour can change materially without any change to the deployer's code.
- Prompt drift: Research has documented behavioural degradation of up to 46% over 500 interactions. An agent certified as compliant today may drift out of compliance by next week.
- Adversarial evolution: Attack techniques evolve faster than quarterly audit cycles. Jailbreaks discovered on Tuesday may not be tested until the next quarterly assessment.
This isn't a criticism of periodic certification — it's a recognition that agents need an additional, faster feedback loop.
Hardware Attestation: The Identity Layer
Yubico's partnership with Delinea (announced March 2026) introduces Role Delegation Tokens — hardware-attested proof that a specific human approved a specific agent action using a physical YubiKey. HID Global is evolving PKI to issue digital certificates for AI agents, addressing dynamic identity lifecycles and capability attestation.
These are Layer 1 solutions: they answer "who is this agent?" and "who authorised it?" Runtime attestation answers the complementary question: "what has this agent been doing since that authorisation?" Hardware identity + runtime behaviour = complete trust picture.
Add Runtime Attestation to Your Trust Stack
Already certified with AIUC-1, ISO 42001, or CSA STAR? AgentApproved proves your agents are living up to that certification every day.