Why ELTON · FDA-Compliant

The model ships
without its pipeline.

FDA names seven cybersecurity threats for AI devices. ELTON tests all seven, from the training data to the deployed device.

SEVEN AI THREATS, FOUR ATTACK SURFACESFDA SECTION XIITraining datadatasets and governance, before a model existsDATA POISONINGMODEL BIASModel and inferencethe registry, the artifacts, and the inference APIMODEL INVERSION / STEALINGDeployed devicecrafted inputs against the running modelMODEL EVASIONthe one surface most programs already testPipeline and postmarketmonitoring and evaluation, after releaseDATA LEAKAGEOVERFITTINGPERFORMANCE DRIFTSix of the seven threats begin upstream of the deployed device.ELTON tests all seven, and maps every test to a named FDA threat.7 OF 7SEVEN AI THREATS, FOUR SURFACESTraining dataDATA POISONINGMODEL BIASModel and inferenceMODEL INVERSION / STEALINGDeployed deviceMODEL EVASIONPipeline and postmarketDATA LEAKAGEOVERFITTINGplus performance drift on the postmarket surfaceSix of seven begin upstream ofthe deployed device.ELTON tests all seven, mapped to each FDA threat.7 OF 7 TESTED
FDA’s January 2025 draft guidance, Section XII, names seven cybersecurity threats specific to AI-enabled devices.
FDA, January 2025 draft guidance

FDA now names
the AI threats.

Section XII lists seven cybersecurity threats that are specific to AI devices.

Who it applies to

Cyber devices under section 524B carry a cybersecurity requirement already: device software, able to connect, with characteristics that could be attacked. For those devices, the draft guidance adds the AI-specific risks their submissions should address.

The seven named threats

Data poisoning, model inversion or stealing, model evasion, data leakage, overfitting, model bias, and performance drift. Six of the seven reach the model through the data and infrastructure around it, not the deployed model alone.

What FDA asks for

Cybersecurity risk management that covers these AI risks, and testing matched to them, on top of the 2023 premarket cybersecurity guidance. The evidence is what an FDA reviewer reads.

The guidance is still a draft. The exact wording can move before it is final, and this page does not treat it as settled. What will not move is the direction: a device whose function depends on a model has an attack surface that starts in the training data and runs through to postmarket monitoring. Test the deployed model only, and you have evidence for one of the seven threats. Why the pipeline is the attack surface →
Our position · Premarket

Test the pipeline
as the product.

Six of the seven threats originate upstream of the deployed device, so the training data, registries, evaluation logic and deployment infrastructure are tested as part of it.

FIRST · PENETRATION TEST OF THE AI/ML PIPELINE ENVIRONMENTTraining serversaccess, jobs, credentialsModel registriesartifacts, versions, accessData storesdatabases, object stores, file systemsDeployment infrastructureserving path, CI/CD, secretsTRAINING DATADATA POISONINGMODEL BIASStores, ingestion, labelingunauthorized reads and writes, injection on ETL,batch integrity, annotation access, provenanceMODEL AND INFERENCEMODEL INVERSION OR STEALINGRegistry, artifacts, inference APIauth and access, excess disclosure, rate limits,enumeration, encryption at rest, serialized filesDEPLOYED DEVICE, RUNTIMEMODEL EVASIONThe inference path on the devicemalformed, corrupted and out-of-distribution inputs,boundary cases, preprocessing bypass, fail-safe, logsPIPELINE AND DEPLOYMENTDATA LEAKAGEWhat leaves the pipelineAPI responses, logs, telemetry and debug endpoints,encryption in transit and at rest, training remnantsTRAINING AND EVALUATIONOVERFITTINGEvaluation and validation infrastructureaccess to evaluation data, scripts and results,split and threshold tampering, duties, versioningPOSTMARKET MONITORINGPERFORMANCE DRIFTDrift detection and retraining controlsthresholds and alert configuration, data collectionendpoints, alert channels, retraining triggersThirty-seven hands-on tests across six attack surfaces, every one tied to a named FDA threat.Reproducible findings with evidence, CVSSv3 and CVSSv4 ratings on the MDDT-qualified rubric, VEX-aligned disposition, one threat-mapped report.TECHNICAL TESTINGFIRST · PENTEST THE PIPELINE ENVIRONMENTTraining serversaccess, jobs, credentialsModel registriesartifacts, versions, accessData storesdatabases, object stores, file systemsDeployment infrastructureserving path, CI/CD, secretsTRAINING DATADATA POISONINGMODEL BIASStores, ingestion, labelingunauthorized reads and writes, injection on ETL,batch integrity, annotation access, provenanceMODEL AND INFERENCEMODEL INVERSION OR STEALINGRegistry, artifacts, inference APIauth and access, excess disclosure, rate limits,enumeration, encryption at rest, serialized filesDEPLOYED DEVICE, RUNTIMEMODEL EVASIONThe inference path on the devicemalformed, corrupted and out-of-distribution inputs,boundary cases, preprocessing bypass, fail-safe, logsPIPELINE AND DEPLOYMENTDATA LEAKAGEWhat leaves the pipelineAPI responses, logs, telemetry and debug endpoints,encryption in transit and at rest, training remnantsTRAINING AND EVALUATIONOVERFITTINGEvaluation and validation infrastructureaccess to evaluation data, scripts and results,split and threshold tampering, duties, versioningPOSTMARKET MONITORINGPERFORMANCE DRIFTDrift detection and retraining controlsthresholds and alert configuration, data collectionendpoints, alert channels, retraining triggersThirty-seven hands-on tests,every one tied to a named FDA threat.Reproducible, evidenced, rated CVSSv3 and v4,dispositioned in the VEX lifecycle.TECHNICAL TESTING
The pipeline environment is tested first. The threat-specific tests follow, surface by surface, and every one maps to a named FDA threat.

Infrastructure first

Testing starts with a penetration test of the AI/ML pipeline environment: training servers, model registries, data stores and deployment infrastructure. The threat-specific activities follow, and they inherit what the infrastructure test found.

Reproducible, with evidence

Every technical test is hands-on penetration testing against pipeline infrastructure, access controls, APIs and the deployed device. Each finding is reproducible and comes with evidence. None of it needs model weights or training data from you.

One of seven, every time

Each test maps to one of the seven threats FDA names in Section XII of its January 2025 draft guidance. The report reads the way a reviewer reads: threat by threat, with the tests, results and evidence under each.

The test plan

Seven threats,
thirty-seven tests.

What ELTON runs on a premarket AI-enabled device, listed so you can read coverage before the first packet is sent.

Before the seven

The pipeline environment
  • A penetration test of the AI/ML pipeline environment: training servers, model registries, data stores, deployment infrastructure. Everything below inherits its findings.

Data poisoning

Training data
  • Penetration test of the training data stores, databases, object stores and file systems, with unauthorized reads and writes attempted.
  • Injection testing on data ingestion endpoints and ETL pipelines.
  • Attempts to write, modify or delete training datasets through every reachable interface.
  • Integrity controls on training data batches: hash chains, checksums, signatures.
  • Access control testing on the labeling and annotation tools.
  • Segmentation between the training environment and the production and corporate networks.

Model inversion or stealing

Model and inference
  • Penetration test of the model registry and artifact storage: authentication, authorization, access controls.
  • The inference API probed for excess disclosure: confidence scores, raw logits, class probabilities, softmax outputs.
  • Rate limiting, authentication and abuse controls on the inference endpoints.
  • Model architecture enumeration through error messages, response timing and debug endpoints.
  • Model file encryption at rest and integrity verification on load.
  • Search for exposed serialization files (pickle, ONNX, SavedModel) through directory traversal or misconfigured storage.

Model evasion

Deployed device, runtime
  • Input validation and sanitization on the deployed inference path: malformed, corrupted and out-of-distribution inputs.
  • Boundary testing with crafted edge cases at the model’s decision boundaries.
  • Device behavior on inputs outside the training distribution: graceful degradation, error handling, a fail-safe response.
  • Bypass testing on the input preprocessing pipeline: normalization, filtering, format conversion.
  • Confirmation that the device logs, alerts on or rejects anomalous inputs.

Data leakage

Pipeline and deployment
  • Inference API responses audited for data the caller does not need: training metadata, internal identifiers, patient-derived features.
  • Application and infrastructure logs reviewed for model parameters, training data references, PII and PHI.
  • Encryption on inter-service traffic inside the pipeline: training to registry, registry to deployment.
  • Encryption at rest for training datasets, evaluation datasets and model artifacts.
  • Debug, monitoring and telemetry endpoints reviewed for information disclosure.
  • Search for training data remnants in deployed artifacts, container images and configuration files.

Overfitting

Training and evaluation
  • Penetration test of the evaluation and validation infrastructure: evaluation datasets, scripts and results.
  • Access control on hyperparameter configurations and training job definitions.
  • Separation of duties across model training, evaluation and approval roles.
  • Attempts to manipulate the training, validation and test splits, substitute an evaluation split or change an acceptance threshold, and whether version control catches it.
  • Audit of version control and change tracking on evaluation pipelines and acceptance criteria.

Model bias

Training data
  • Access control on labeling and annotation workflows against unauthorized label changes.
  • Audit of provenance tracking and chain of custody on training datasets.
  • Governance controls that prevent unauthorized changes to dataset composition or demographic distribution.
  • Audit logging on training data selection, filtering and sampling decisions.

Performance drift

Postmarket monitoring
  • Penetration test of the performance monitoring and drift detection infrastructure.
  • Attempts to change drift thresholds, alert configuration and monitoring parameters without authorization.
  • Integrity and authentication on postmarket data collection endpoints.
  • Alert and notification channels (email, webhook, dashboard) tested for spoofing and suppression.
  • Access controls on model update and retraining triggers.
Beyond the technical tests

When the access exists,
the attack runs.

Five adversarial exercises, each bounded to one target, one batch or one feed, run when you can provide the model, data or environment each one needs.

Adversarial inputs

Working evasion inputs against representative inputs: gradient methods (PGD, FGSM, C&W) with white-box access, transfer or query methods without it. The perturbation matches the input modality, and the exercise needs your preprocessing pipeline, or it does not run.

A poisoned batch, through your controls

A deliberately poisoned or shifted batch run through the data-validation controls, recording whether outlier, activation-clustering or spectral checks flag it, and how strongly. Needs model access for activations and a dataset ELTON may poison outside production.

Membership inference

A threshold attack on loss or confidence that separates records known to be in the training set from records known to be out. Needs the train and holdout split and outputs that expose confidence or loss.

Single-record inversion

With white-box access, a gradient-ascent inversion against one target record agreed with you, scored on the reconstruction. GAN-assisted inversion is a separate engagement.

Sub-threshold drift

A constructed data feed with incremental covariate shift calibrated to stay under the detector’s thresholds, in a test or staging environment, measuring how far performance walks before an alert fires.

The rules

Prerequisites are listed up front and confirmed at kickoff. Where a control does not exist, no ingest monitoring, no drift detection, the missing control is the finding and the dependent exercise does not run. Full-scale demonstrations through the training harness, retraining, GAN-assisted inversion and side-channel timing are quoted separately.

Our position · Postmarket

Surveillance has
two halves.

FDA’s postmarket cybersecurity expectations cover the device, and IEC PAS 63621:2026 covers the data behind its model. An AI-enabled device needs both, for its whole life.

THE MANUFACTURER'S DATA LIFECYCLE · IEC PAS 63621:2026Requirementscriteria and targetsPlanningcoverage of intended useAcquisitionsources, criteria, DOUPDevelopmentannotation, verificationProvisioningsealed verification setDecommissioningremoval and retentionFIELD DATAPostmarket collectionendpoints, consent, PII removalQUALITY VERIFICATIONAgainst acceptance criteriaon a risk-based schedule (6.5.6)ANALYSISRoot cause, action, monitoringdrift, or tampering? (6.5.7)RETRAIN AND RELEASEA new release in the twinnew dataset version, sealed splitTHE LOOP RUNS FOR THE LIFE OF THE DEVICEWHAT ELTON TESTS UNDERNEATH IT, RELEASE AFTER RELEASECollection endpointsintegrity, authentication, PII pathsThe verification recordversions, immutable IDs, provenanceDrift detectionthresholds, alerts, suppressionRetraining and releaseregistry, artifacts, release pathSurveillance you can tamper with is not surveillance.The data record belongs to the manufacturer under IEC PAS 63621. Whether it can be trusted is what ELTON tests, release after release.FDA 524B · IEC PAS 63621DATA LIFECYCLE · IEC PAS 63621:2026Requirementscriteria and targetsPlanningcoverage of intended useAcquisitionsources, criteria, DOUPDevelopmentannotation, verificationProvisioningsealed verification setDecommissioningremoval and retentionIN THE FIELD · THE LOOPPostmarket collectionendpoints, consent, PII removalAgainst acceptance criteriaon a risk-based schedule (6.5.6)Root cause, action, monitoringdrift, or tampering? (6.5.7)A new release in the twinnew dataset version, sealed splitWHAT ELTON TESTS UNDERNEATH ITCollection endpointsintegrity, authentication, PII pathsThe verification recordversions, immutable IDs, provenanceDrift detectionthresholds, alerts, suppressionRetraining and releaseregistry, artifacts, release pathSurveillance you can tamper withis not surveillance.The data record is the manufacturer's.Whether it can be trusted is what ELTON tests.FDA 524B · IEC PAS 63621
The lifecycle clauses of IEC PAS 63621 along the top, the field loop in the middle, and the controls ELTON tests so the loop can be trusted.

Data quality is a lifecycle activity

IEC PAS 63621 is explicit: data that met requirements at verification can stop meeting them later, through drift, hardware and software changes, or activity inside or outside the product (6.5.7). Verification is repeated on a schedule set by the data quality risk, and every issue gets a root cause, a corrective action and further monitoring. That record is the manufacturer’s. ELTON’s job is to make sure it can be trusted.

Test the surveillance controls, every release

The performance drift tests in the premarket plan do not retire at clearance. Drift detection infrastructure, thresholds and alert configuration, postmarket data collection endpoints, alert channels and retraining triggers are tested again with each release and at least annually, because a detector an attacker can quiet is the first thing an attacker will quiet.

Every retrain is a release

A model update is a new release in the digital twin: new dataset version identifiers, provenance for every label, a verification set the training phase could not see (6.6), and a rerun of the threat tests on what changed. Findings and their VEX status carry across releases, so the record reads as one product rather than a stack of reports.

The record has to hold up

The specification asks for revision control with integrity protection, unique and immutable data identifiers, annotator identity, documented storage security and permissions, de-identification, and a plan for removing data when a device leaves clinical use (6.7). Each one is a control an attacker would target. ELTON tests the integrity of dataset versions, the access paths to annotation tooling and storage, where PII and PHI land in logs and artifacts, and the data removal paths, alongside the SBOM and CVE monitoring that runs on the device itself.

Standards and evidence

Evidence mapped
to every threat.

One set of evidence that travels across the frameworks arriving now.

Threat-mapped report

An AI-Enabled Device Cybersecurity Assessment Report. Every test result and bounded exercise tied to one of the seven named threats, with traceable evidence, ready to carry into the submission.

Ratings that hold up

Findings carry CVSSv3 and CVSSv4 ratings where they apply, through the MDDT-qualified rubric. The rating is product-adjusted and evidenced, not a generic score.

Status you can manage

Every finding runs the VEX-aligned lifecycle on the ELTON platform, alongside the rest of the product’s findings, and carries across releases.

ELTON meets FDA premarket (Section 524B) and postmarket cybersecurity vulnerability testing and management requirements. The AI threats sit inside that work. They also map onto IEC PAS 63621, the 2026 data-management specification for AI-enabled medical devices, and onto the EU and other regimes now adding AI expectations. The rules are arriving on different timelines, so evidence mapped to named threats travels better than a separate test report for each.
Questions

Questions about
AI device testing.

Does FDA require AI-specific cybersecurity testing?

For a cyber device under section 524B, cybersecurity is already a requirement. FDA’s January 2025 draft guidance for AI-enabled device software functions then names seven AI-specific threats, in Section XII, and recommends that sponsors address them in cybersecurity risk management and testing. The guidance is still draft, so ELTON treats it as direction, not settled text, and ties the hard requirement to 524B.

What are the seven threats?

Data poisoning, model inversion or stealing, model evasion, data leakage, overfitting, model bias, and performance drift. Only model evasion happens on the deployed model at runtime. The other six begin in the training data, the model and inference layer, the pipeline, or postmarket monitoring.

Do you need access to our model and training data?

The thirty-seven technical tests do not. They run against pipeline infrastructure, access controls, APIs and the deployed device. The bounded adversarial exercises do, and each lists exactly what it needs up front: model weights or gradients, a dataset ELTON may poison outside production, the train and holdout split, detector thresholds. If you cannot provide a prerequisite, ELTON assesses that threat from the architecture and the technical findings and records the gap, so the engagement still completes.

What is the difference between a technical test and a bounded exercise?

A technical test is hands-on penetration testing that produces a reproducible finding with evidence. A bounded exercise runs an advanced adversarial technique against the model or its pipeline at a fixed size, one target, one batch, one feed, and scores the result, without a research-grade campaign.

What does proper postmarket surveillance look like for an AI-enabled device?

Two halves, both continuous. The cybersecurity half is what FDA expects of any cyber device after release: software and hardware materials monitored against global feeds, new CVEs correlated with test results, exploitability verified, findings dispositioned. The data half comes from IEC PAS 63621:2026: data quality verified on a schedule set by risk, drift analyzed to a root cause, every retrain handled as a release with a sealed verification set. ELTON runs the first half on the device and its pipeline, and tests the controls the second half depends on.

What is IEC PAS 63621, and does it apply to us?

A Publicly Available Specification from IEC technical committee 62, published March 2026, setting a framework for the lifecycle of data used to train, test or validate an AI model that is part of a medical device: requirements, planning, acquisition, development, provisioning and decommissioning. It is a specification, not a regulation. It is also the most specific description yet of what a defensible data record looks like, and the controls it asks for, revision control with integrity protection, immutable identifiers, provenance, sealed verification data, are the ones ELTON tests.

Is this the same as the AI you use to find vulnerabilities?

No. The pipeline that finds vulnerabilities uses AI to build tools and decide which to run; your data never enters a prompt. AI-Enabled Device Testing is about the AI inside your device, and the seven ways it can be attacked. A pipeline, not a prompt, on our side; the model and its pipeline, on yours.

Does it matter where the model and its pipeline run?

No. The tests follow the pipeline wherever it lives: your data center, your cloud tenant, or a vendor’s managed ML platform. What changes is the access and the prerequisites, agreed at kickoff, and the digital twin records each component where it actually runs.

How does this map to our FDA submission?

The report maps every result to one of the seven named threats, with evidence, and findings carry CVSSv3 and CVSSv4 ratings through the MDDT-qualified rubric and a VEX-aligned status. A reviewer reads the AI-specific risks addressed, tested, rated and dispositioned, inside the 524B cybersecurity record.

Exploitability management for medical devices. FDA §524B methodologyExploitability proven at runtime95% faster than legacy testing Book a Demo →
Platform
OverviewAvoid FDA DeficienciesAvoid Consulting FeesDigital Twin TraceabilityAI MedDevice PentestingFDA Vulnerability VerificationVulnerability ChainingRemediation OptimizationRemote TestLink™Incident ResponseAutomated VEX & MetricsCVSSv4 Migration
Solutions
EnterpriseStartups / SMBs Postmarket SurveillanceIncident ResponseSecurity EngineeringRegulatory AffairsFDA §524BEU MDR/CRAEU REDNIS2IMDRF N60 / N73Japan MHLWAI-Enabled DevicesDevice Modalities
Why ELTON
100% ComplianceFDA TraceabilityFDA Vulnerability VerificationFDA-Qualified CVSSv3CVSSv4 MigrationQMSR Audit Compliance One SolutionTesting ProcessPosture ManagementAI-PoweredELTON vs. Legacy TestingPricing
Resources
FDA Deficiency ListFDA Testing RequirementsFDA Cyber SOPs & TemplatesVulnerability Disclosure TemplatesExample Test ReportRemediation LibraryRegulatory GuidesWebinarsAI NewsletterThe End of Legacy TestingThe AI Vulnerability ExplosionAI Inside the ProductCybersecurity TestingSecurity AdvisoriesWhitepapersIntelligence & Blog
Company
AboutCredentialsLeadershipPartnershipsData SecurityContact Meet ELTON →
See the graph decide, live →>