PricingDemo
Log InGet API Key
AI Governance Infrastructure — Cryptographic Enforcement

AI Guardrails That Prove,
Not Just Filter

Most AI guardrails are advisory. They filter, log, and hope. Cryptographic guardrails are different. Every AI action becomes a post-quantum signed receipt. Every scope boundary becomes an enforceable constraint. Every governance decision becomes independently replayable. The guardrail is not a suggestion. It is a mathematical proof.

74 B
Per Attestation (H33-74)
3
PQ Hardness Assumptions
42 µs
Per Governance Attestation
Replay Integrity

Most AI Guardrails Are Theater

The AI industry has adopted the word "guardrails" to describe a collection of techniques that share one critical property: none of them produce cryptographic evidence that governance was enforced. Understanding what current guardrails actually do -- and what they cannot do -- is the starting point for understanding why cryptographic enforcement matters.

Prompt-Based Guardrails

The most common form of AI guardrail is a system prompt that instructs the model to avoid certain behaviors. "Do not discuss competitors." "Do not generate harmful content." "Do not execute financial transactions above $10,000." These instructions exist in natural language, are interpreted by the model at inference time, and can be circumvented through prompt injection, jailbreaking, or simple model confusion. There is no cryptographic record that the instruction existed. There is no proof that the model followed it. There is no way for an independent third party to verify, after the fact, that the guardrail was active during a specific interaction.

Content Filtering

A step above prompt-based guardrails, content filters examine model inputs and outputs against classification rules. Toxicity filters. PII detection. Topic blocking. These systems operate as middleware: they intercept traffic and make allow/deny decisions based on pattern matching or secondary model classification. The problem is not that these filters do not work. Many of them work well for their narrow purpose. The problem is that they produce no verifiable evidence of their operation. A content filter that blocks a request generates a log entry in a database controlled by the platform operator. That log entry can be modified, deleted, or fabricated. No independent party can verify that the filter was active, that its rules matched the claimed policy, or that the decision was correct.

Rate Limiting and Usage Caps

Enterprise AI deployments commonly implement rate limits, spending caps, and usage quotas as governance mechanisms. An agent can make no more than 100 API calls per hour. A department's AI budget cannot exceed $50,000 per month. These are operational constraints, not governance proofs. They exist in the platform's accounting system. They can be retroactively adjusted. They produce no independent evidence that the constraint was active at any specific point in time. When an insurer asks whether spending caps were enforced during the period covered by a claim, the organization can produce spreadsheets and dashboards. It cannot produce cryptographic proof.

Logging and Audit Trails

The most sophisticated current approach to AI governance is comprehensive logging. Every model call, every agent action, every tool invocation is recorded in a centralized logging system. The logs are then available for audit. This sounds reasonable until you consider the trust model. The entity that produces the logs is the same entity that controls the logging infrastructure. Logs can be selectively deleted. Timestamps can be retroactively modified. Log entries can be fabricated. The organization being audited is also the organization that controls the evidence. This is not a governance system. It is a self-certification system.

The Common Thread

Every current approach to AI guardrails shares the same structural weakness: they produce no evidence that can be independently verified by a party that does not trust the platform operator. Prompt instructions cannot be proven to have existed. Filter decisions cannot be replayed. Rate limits cannot be independently confirmed. Logs cannot be distinguished from fabricated records. This is not a minor gap. It is the gap. When regulators require AI governance, when insurers underwrite AI risk, when enterprises deploy autonomous agents -- the question is not whether guardrails exist. The question is whether anyone can prove that guardrails were enforced.

What Cryptographic Guardrails Mean

A cryptographic guardrail is not a filter, a prompt, or a log entry. It is an attestation. Every AI action -- every model call, every agent decision, every tool invocation, every scope boundary check -- produces a post-quantum signed receipt that captures the complete governance context at execution time. The receipt is not a description of what happened. It is a mathematical proof that specific governance constraints were active and that the action fell within the boundaries defined by those constraints.

Attestation, Not Permission

Traditional guardrails operate on a permission model. The system checks whether an action is allowed and either permits or blocks it. The decision lives in memory and may or may not be logged. Cryptographic guardrails operate on an attestation model. Every action -- permitted or denied -- produces a signed attestation that captures the governance graph state, the policy hash, the scope boundaries, the signer set, and the action itself. The attestation is immutable once created. It is independently verifiable. It is deterministically replayable. The difference is fundamental. A permission system answers "was this allowed?" A cryptographic attestation system answers "can anyone, anywhere, at any future time, independently verify what was allowed, what happened, and whether the action was within governance bounds?"

The Six Fields of Every Guardrail Attestation

Every AI guardrail attestation in H33 captures six fields that together constitute a complete governance record:

These six fields are not optional metadata. They are the guardrail. If any field is absent, the attestation is invalid. If any field is modified after creation, the hash chain breaks and every downstream attestation becomes detectably tampered. This structure is described in detail in the AI Decision Attestation specification.

Post-Quantum Signature Security

Every guardrail attestation is signed using three independent post-quantum signature families: ML-DSA (lattice-based, NIST FIPS 204), FALCON (NTRU lattice-based), and SLH-DSA (stateless hash-based). Breaking a single attestation requires simultaneously defeating MLWE lattices, NTRU lattices, and stateless hash functions -- three independent hardness assumptions. The resulting signature bundle is compressed to 74 bytes via H33-74 while preserving full independent verifiability. This is not theoretical future-proofing. It is production cryptography protecting every AI governance decision today.

What Every AI Action Proves

The value of cryptographic guardrails is measured by what they prove. Not what they claim. Not what they log. What they prove, to an independent verifier, without requiring trust in any platform operator.

Scope Enforcement: Agents Provably Cannot Exceed Authority

When an AI agent operates under H33 governance, its scope is defined by a governance graph -- a directed acyclic structure where each node represents an authority boundary and each edge represents a delegation. The agent's scope is the intersection of all delegations from the root authority to its position in the graph. This is not a configuration setting. It is a cryptographic structure where every delegation is signed and every scope boundary is attested. An agent cannot exceed its authority because exceeding authority would require producing a valid attestation for an action outside its scope -- and valid attestations require signatures from authority nodes that did not delegate that capability. The agent governance architecture enforces this at the infrastructure layer, below the model, below the application, below any code that the agent itself can influence.

Negative Authority Proofs: Prove What AI Cannot Do

Most governance systems can only prove what was permitted. H33 also proves what was not permitted. A negative authority proof is a cryptographic attestation that a specific capability was explicitly excluded from an agent's scope at a specific time. If an agent's governance graph does not include financial transaction authority, H33 produces a signed proof of that exclusion -- not merely the absence of a permission, but the presence of a cryptographic proof that the permission does not exist. This is critical for insurance underwriting. When an insurer needs to know that an AI agent could not have initiated wire transfers during a specific period, a negative authority proof provides independently verifiable evidence of that constraint. The alternative -- searching logs for the absence of wire transfer activity -- proves nothing about capability, only about observed behavior.

Deterministic Replay: Reconstruct Any AI Decision

Every guardrail attestation includes sufficient context to deterministically replay the governance decision. Given the governance graph state, the policy hash, the scope boundaries, and the action parameters, any independent verifier implementation can reconstruct the identical attestation chain and arrive at the identical governance verdict. This is not log review. It is mathematical reconstruction. The same inputs always produce the same outputs. Across implementations. Across languages. Across time. The agent replay system provides the tooling for this reconstruction, and the governance replay demo provides a live demonstration of deterministic replay in action.

Tamper Detection: Modify One Decision, Break the Chain

Guardrail attestations are hash-chained. Each attestation includes the hash of the previous attestation in the chain. Modifying a single attestation -- changing an action, altering a scope boundary, backdating a timestamp -- changes its hash, which invalidates every subsequent attestation in the chain. An attacker cannot selectively modify history without detection. The chain either verifies completely or it does not. There is no partial tampering.

Independent Verification: No Vendor Trust Required

Guardrail attestations are verified using public verifier implementations. The verifier does not require access to H33 infrastructure. It does not require an API key. It does not require network connectivity. An air-gapped machine running an independent verifier implementation can validate any attestation chain. The HATS protocol defines the verification semantics, and any conformant implementation produces identical results. This is what separates cryptographic guardrails from every alternative: the evidence is not controlled by the entity being evaluated. Read more about verifiable AI actions and the HATS standard that governs verification semantics.

Comparison

Traditional AI Guardrails vs Cryptographic Enforcement

A structural comparison of what current AI guardrail approaches provide versus what cryptographic enforcement delivers.

DimensionTraditional AI GuardrailsH33 Cryptographic Enforcement
Evidence typeLog entries, dashboard metrics, self-reported compliancePost-quantum signed attestation receipts, hash-chained
Tamper resistanceDatabase access controls (platform-controlled)Hash chain integrity -- modify one record, break entire chain
Independent verificationRequires platform access and platform trustOffline verification with public verifier, no vendor trust
Scope enforcementApplication-layer permission checksCryptographic governance graph with signed delegations
Negative proofsNot possible -- absence of logs proves nothingSigned negative authority proofs for excluded capabilities
Replay capabilityLog review (non-deterministic, platform-dependent)Deterministic replay producing identical outputs anywhere
Quantum resistanceNone (RSA/ECDSA signatures vulnerable to quantum attack)Three independent PQ hardness assumptions per attestation
Regulatory evidenceQuestionnaire responses, audit interviews, screenshotsMachine-verifiable conformance evidence, HATS-compliant
Insurance valueSelf-reported risk posture assessmentsIndependently verifiable governance proofs for claim adjudication
Time to verifyWeeks to months (manual audit cycles)Milliseconds (automated cryptographic verification)

EU AI Act and Regulatory Readiness

The EU AI Act, effective August 2026, imposes specific obligations on providers and deployers of high-risk AI systems. These obligations require capabilities that traditional AI guardrails cannot provide. Cryptographic guardrails map directly to the Act's requirements.

Article 9 -- Risk Management: The Act requires a "risk management system" that operates throughout the AI system's lifecycle. H33 provides continuous governance attestation -- not periodic risk assessments, but per-action cryptographic proof that risk management constraints were enforced. Every action is attested. Every scope boundary is enforced. Every policy change is captured in the hash chain.

Article 12 -- Record-Keeping: High-risk AI systems must enable "automatic recording of events" with sufficient detail for conformity assessment. H33 attestation chains are automatic, tamper-evident, and deterministically replayable. They do not require manual configuration. They cannot be selectively disabled. They provide exactly the record-keeping granularity that conformity assessment demands.

Article 14 -- Human Oversight: The Act requires human oversight mechanisms that enable intervention and override. H33's governance graph architecture supports human-in-the-loop attestation -- specific authority nodes can require human approval before delegating execution authority, and that approval is captured as a signed attestation in the chain.

Article 17 -- Quality Management: Providers must implement quality management systems including "examination, test and validation procedures." H33's continuous governance framework and the HATS conformance protocol provide exactly this: machine-verifiable test procedures with deterministic expected outputs and continuous control monitoring that operates at the cryptographic layer.

For a complete mapping of H33 capabilities to regulatory frameworks including the NIST AI Risk Management Framework, see AI Compliance Infrastructure.

How Cryptographic Guardrails Work

Cryptographic guardrails operate at three layers: the governance graph layer, the attestation layer, and the verification layer. Each layer is independent. Each layer produces independently verifiable outputs. Together, they form a complete governance infrastructure that replaces advisory controls with mathematical proof.

The Governance Graph

Every AI deployment under H33 governance is described by a directed acyclic graph (DAG) where nodes represent authority boundaries and edges represent signed delegations. The root authority defines the maximum scope of the deployment. Each delegation narrows the scope -- an agent at depth 3 in the graph can only exercise the intersection of all delegations from root to its node. The graph is versioned. Every change produces a new graph hash. Graph state is captured in every attestation, binding each AI action to the specific authority structure that governed it.

The Attestation Pipeline

When an AI action occurs -- a model call, a tool invocation, a scope boundary check -- the attestation pipeline captures the six governance fields (action, authority, scope, timestamp, policy hash, signer set), constructs the attestation, chains it to the previous attestation hash, and signs it with the three-family post-quantum signature bundle. The attestation is then compressed to 74 bytes via H33-74. The entire pipeline executes in 42 microseconds. There is no perceptible latency impact on AI operations.

The Verification Layer

Verification is independent of attestation. Any conformant HATS verifier implementation can validate an attestation chain without access to H33 infrastructure. The verifier checks signature validity across all three PQ families, reconstructs the hash chain, verifies governance graph consistency, and produces a deterministic verification verdict. The same attestation chain always produces the same verdict, regardless of which verifier implementation performs the check.

$ h33 verify governance-chain ./attestations/

Chain length: 2,847 attestations
Time range: 2026-05-01T00:00:00Z to 2026-05-18T23:59:59Z
Hash chain: VALID (continuous, no gaps)
Signatures: VALID (ML-DSA + FALCON + SLH-DSA)
Scope violations: 0
Negative proofs: 147 (all valid)
Verdict: CONFORMANT

Deployment Patterns for AI Guardrails

Pattern 1: Agent Governance Wrapper

The most common deployment pattern wraps existing AI agents with H33 governance attestation. The agent operates normally -- making model calls, invoking tools, processing data -- while the H33 governance layer intercepts each action, validates it against the governance graph, produces the attestation, and chains it to the previous record. The agent code does not change. The governance enforcement operates below the application layer. This pattern is described in detail in the AI Agent Governance documentation.

Pattern 2: Model Gateway Attestation

For organizations that route AI traffic through API gateways, H33 integrates at the gateway layer. Every model call through the gateway produces a guardrail attestation capturing the requesting identity, the model version, the policy in effect, and the scope boundaries. This pattern provides governance coverage for all AI traffic regardless of which application or agent originates the request.

Pattern 3: Continuous Governance Monitoring

Beyond per-action attestation, H33 provides continuous governance monitoring that attests the state of the entire AI deployment at regular intervals. Model versions, policy configurations, scope boundaries, delegation structures -- all are captured in periodic governance state attestations. This provides a cryptographic timeline of the governance posture, independent of individual action attestations. See AI Operational Integrity for the continuous monitoring architecture.

Pattern 4: Insurance-Grade Evidence Collection

For organizations that need to demonstrate AI governance to insurers, H33 produces attestation bundles specifically structured for claim verification. These bundles include the governance graph, the attestation chain, the negative authority proofs, and the deterministic replay instructions -- everything an insurer needs to independently verify that governance was enforced during the period covered by a claim.

Frequently Asked Questions

AI Guardrails FAQ

What are cryptographic AI guardrails?
Cryptographic AI guardrails replace advisory filters and logging with post-quantum signed attestation receipts for every AI action. Each decision, scope boundary, and authority delegation is captured as a tamper-evident, independently verifiable cryptographic record. The guardrail is not a suggestion the AI can override. It is a mathematical proof that specific governance constraints were active at execution time.
How do cryptographic guardrails differ from prompt-based AI safety?
Prompt-based guardrails instruct the AI to avoid certain behaviors. The AI can ignore, misinterpret, or be manipulated past those instructions. Cryptographic guardrails operate at the infrastructure layer: every action is signed with the governing policy hash, scope boundaries are enforced before execution reaches the model, and the resulting attestation chain is independently replayable. The difference is between asking and proving.
Can cryptographic guardrails prove what an AI agent cannot do?
Yes. H33 implements negative authority proofs -- cryptographic attestations that a specific capability was not granted at a specific time. If an agent's scope excludes financial transactions, the governance graph contains a signed negative proof for that capability class. This is not the absence of a permission. It is the presence of a cryptographic proof that the permission does not exist.
How does deterministic replay work for AI governance?
Every AI action attestation includes the governance graph state, policy hashes, scope boundaries, and signer set active at execution time. Given the same inputs and governance state, any independent verifier implementation can reconstruct the identical attestation chain and arrive at the identical governance verdict. This is not log review. It is mathematical reconstruction of the governance state that governed a specific AI action.
Are H33 AI guardrails compliant with the EU AI Act?
The EU AI Act requires high-risk AI systems to maintain logs, implement human oversight, and demonstrate conformity assessment. H33 cryptographic guardrails provide the evidence infrastructure these requirements demand: tamper-evident action records, scope-enforced authority boundaries, deterministic replay for conformity assessment, and independently verifiable governance proofs. H33 does not replace legal compliance. It provides the cryptographic evidence that compliance claims can be verified against.
What post-quantum algorithms protect AI guardrail attestations?
H33 attestations are signed using three independent post-quantum signature families: ML-DSA (lattice-based, NIST FIPS 204), FALCON (NTRU lattice-based), and SLH-DSA (stateless hash-based). An attacker would need to simultaneously break MLWE lattices, NTRU lattices, and stateless hash functions to forge a single guardrail attestation. Each attestation is compressed to 74 bytes via H33-74 while preserving full independent verifiability.

Watch Agent Replay Demo

See cryptographic guardrails in action. Watch an AI agent operate under governance constraints, then replay the entire decision chain independently.

Watch Agent Replay Demo Agent Governance Architecture