Advisory · AIUC-1 Readiness · Assurance · Audit-Ready
Get your AI agent AIUC-1 certified.
AIUC-1 is the certification standard built for AI agents — security, safety and reliability, independently audited and tested every quarter. Prompt Shields runs the gap analysis, hardens the controls, assembles the evidence and coordinates the accredited third-party audit, so you walk in audit-ready and walk out certified.
Background · The Standard
A certification built for AI that acts, not just answers.
AIUC-1 is the industry standard for AI agent security, safety and reliability. Traditional AI chatbots produce text for a human to read. Agents produce executable actions — code, API calls, database writes, deployed applications — that touch production systems and real data. A vulnerability there isn't hypothetical; it's a live exposure.
The standard is organised as six core principles and 51 requirements, each with one to five controls and a defined evidence model. It is maintained quarterly by a consortium of enterprise CISOs, academic researchers and frontier AI platform builders — and, crucially, it is independently audited and red-teamed, not self-attested.
6
core principles
51
requirements
4
evidence categories
Quarterly
third-party red-teaming
Why It Exists
The frameworks you already have leave a gap.
Enterprise buyers evaluating an AI product ask for evidence — and the standards they lean on were built for different threat models. Each leaves a blind spot that AIUC-1 was designed to close.
Infrastructure assurance, not model assurance.
Encryption, access controls and change management — but a SOC 2-attested product can still generate systematically insecure code or leak credentials through its outputs.
Governance layer, not control layer.
Establishes an AI management system, but doesn't define what the risks are for a specific modality or prescribe the technical controls to address them.
A law, not a technical standard.
Sets what high-risk systems must achieve — critical for legal compliance — without specifying the controls needed to get there.
A risk catalog, not a certification.
A valuable taxonomy of agentic vulnerabilities, but no defined evidence, controls or testing cadence — and no auditor to verify you addressed them.
What's been missing is a framework that is AI-native, modality-specific, technically prescriptive, and independently verifiable through third-party audit and quarterly testing. That's what AIUC-1 provides — and what we get you ready for.
The Standard
Six principles · one standard
Every one of the 51 requirements rolls up to one of these six principles. We assess your product against all of them, and each is where we help you build and prove real controls.
Security
Defend against adversarial attacks, jailbreaks and prompt injection — the input surface for an agent is vast and untrusted.
Safety
Mitigate harmful, biased or out-of-scope output, and protect brand reputation when the agent acts on its own.
Reliability
Prevent hallucinations — including code-specific failures like recommending nonexistent or typosquatted packages.
Accountability
Clear governance and oversight: shared-responsibility models, execution-chain logging and a documented control owner.
Society
Prevent broader harms and align the system with obligations that reach beyond the immediate user.
Data & Privacy
Protect against data leakage, cross-tenant exposure, IP theft and unauthorised use of confidential context.
What Makes It More Than A Checklist
Four evidence categories — shown, not claimed
Each control specifies not just what to do, but what to show and to whom. That evidence model is the hard part of certification — and the part we do the heavy lifting on.
Policy evidence
Terms of service, privacy policies, data-processing agreements and acceptable-use policies — where commitments like “we do not train on customer data” are actually made.
Technical evidence
The real implementation: system-prompt guidance, filtering logic, IAM configs, sandbox policies, credential-detection code and logging. Where an auditor reviews code, not slides.
Operational evidence
Change management, incident response, quarterly internal reviews and vendor due diligence — proof the controls are actively maintained, not just present.
Third-party red-teaming
At least quarterly, across adversarial robustness, harmful outputs, hallucinations and tool-call safety. The layer that prevents self-certification.
How We Help
From first gap analysis to certification recommendation
AIUC-1 is issued by an accredited third-party auditor — nobody self-certifies, and neither do we. Our job is to get you audit-ready and stay beside you through the audit itself. Six ways we do that.
Gap analysis against all six principles
We map your product against the 51 requirements and their controls, mark each Core control as met, partial or missing, and hand you a prioritised, evidence-backed remediation backlog.
Secure-by-default control implementation
Secrets detection on inputs and outputs, tenant-scoped credential isolation, sandbox and egress policy, safe dependency specification against slopsquatting, and least-privilege tool scoping.
Evidence package assembly
We build the policy, technical and operational evidence an accredited auditor expects — organised the way AIUC-1 structures it, so review is fast and there are no surprises.
Quarterly third-party red-teaming
Adversarial robustness, harmful and out-of-scope output, hallucination and tool-call testing — run with our vetted offensive-security partners and delivered as auditor-ready reports.
Runtime governance & logging
Execution-chain logging, approval events and monitoring that satisfy the accountability and enterprise-governance requirements — wired into Atlas AI where it fits your stack.
Audit liaison & coordination
We coordinate the accredited third-party audit end to end, sit in the review with your team, and manage remediation of any findings through to a certification recommendation.
Enterprise buyers no longer accept “trust us.” They ask for the controls, the test results and the audit report. Certification is how you make that conversation short.
How An Engagement Runs
Four phases to audit-ready
We calibrate to where you already are. Teams with strong security foundations move fast; those starting cold get a longer remediation runway. We scope it honestly on the first call.
Scope & gap analysis
We map your AI surface and assess it against the six principles and 51 requirements, producing a scored readiness baseline and a prioritised backlog.
Remediate & implement
We work with your engineers to close Core control gaps — secure defaults, isolation, sandboxing, logging — and stand up the operational processes the standard expects.
Evidence & red-team
We assemble the four evidence categories and run the first quarterly third-party red-team cycle, producing the structured reports the auditor requires.
Audit & certify
We coordinate the accredited third-party audit, support the review, and drive remediation of findings through to a certification recommendation.
What You Walk Away With
Everything an auditor asks for, ready to hand over
Readiness scorecard
Every AIUC-1 requirement rated met / partial / missing, with the Core-vs-Supplemental split called out and gaps ranked by effort and risk.
Control implementation plan
A sequenced, engineer-ready roadmap of the technical and operational controls to build, with quick wins and hard dependencies flagged.
Audit evidence package
Policy, technical, operational and red-teaming evidence assembled and cross-referenced to the requirements, ready to hand to an accredited auditor.
Board-ready assurance summary
A one-page narrative leadership and customers can act on — where you stand against the standard and what certification unlocks commercially.
See where you stand against AIUC-1.
Book a 30-minute readiness call. We'll walk your team through the six principles, sketch where your product likely sits today, and map the shortest honest path to an accredited audit.
Book a readiness call