FDA names seven cybersecurity threats for AI devices. ELTON tests all seven, from the training data to the deployed device.
Section XII lists seven cybersecurity threats that are specific to AI devices.
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 For those devices, the draft guidance adds the AI-specific risks their submissions should address.
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.
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.
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.
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.
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.
Each test maps to one of the seven threats FDA names in Section XII of its January 2025 draft guidance.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.
What ELTON runs on a premarket AI-enabled device, listed so you can read coverage before the first packet is sent.
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.
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 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.
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.
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.
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.
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.
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.
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.
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.
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 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.
One set of evidence that travels across the frameworks arriving now.
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.
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.
Every finding runs the VEX-aligned lifecycle on the ELTON platform, alongside the rest of the product’s findings, and carries across releases.
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.