Problem Space · 01 · Discovery

The vulnerability discovery problem AI created

AI has collapsed the cost of vulnerability discovery to nearly zero, and medical device manufacturers now face more findings per product, per release than any team can triage by hand, each requiring an evidenced disposition under FDA premarket (524B) and postmarket expectations.

This is the trend and the AI-based response: developing your own AI automated analysis to determine which vulnerabilities are actually exploitable against a specific device rather than theoretically present in it. What follows is a practical migration path from manual, point-in-time testing to continuous AI-driven discovery and verification, including how to build a feedback loop that routes verified findings directly into your CI/CD pipeline in a quality-system-compliant manner.

You'll leave with a transition strategy: where to start, what to automate first, and how to keep pace with discovery that no longer has a barrier to entry.

Published CVEs per year, historical and projected. 2024 was the largest single-year jump in CVE history. AI discovery tooling bends the curve from linear to exponential.
172%
Growth since 2020
18K to 40K+ in four years
38%
2024 YoY increase
Largest jump ever recorded
125K+
Projected by 2028
3x current volume
8.5x
Growth 2017 to 2028
AI-driven exponential curve
Critical impact

LLM-powered fuzzers

Intelligent test inputs find edge cases human fuzzers miss. Orders of magnitude more coverage per hour.

Critical impact

AI code auditors

Scan entire codebases for vulnerability patterns in minutes. A junior developer now finds what once took deep expertise.

Critical impact

AI variant finders

One published CVE spawns dozens as the same pattern is matched across every related library and fork.

The pointDiscovery is solved, to the point of drowning. The scarce skill is no longer finding vulnerabilities. It is proving, per device and per release, which ones matter.
Problem Space · 02 · Process

Manual product security is breaking in slow motion

Watch what a product security engineer actually does all day. Five feeds of findings arrive scored five different ways, and every disposition decision lands in a spreadsheet.

The SBOM CVE feed never stops. SAST dumps arrive loaded with false positives. Vulnerability scans, penetration test reports, and third-party research reports each show up with their own scoring assumptions. The engineer modifies the CVSSv3 vector by hand, attempts a CVSSv4 calculation nobody has standardized, and reconciles the result against what a consultant wrote six months ago.

The output is immense variability across sources. The same finding rates 7.8 from the scanner, 5.1 from the pentest report, and 9.1 from the researcher. Bob and Joe triage with different criteria and both are defensible to no one. And all of it, every rating change, every rationale, every disposition, lives in Excel.

FINDING SOURCES · CONSTANT FLOW SBOM CVE feednew matches daily, never stops SAST outputheavy false positive load Vulnerability scansnetwork and host, per release Pentest reportspoint in time, PDF, stale fast Third-party researchresearcher and HDO disclosures ONE ENGINEER Manual re-rating edit CVSSv3 · attempt CVSSv4 CVE-…7.8?open CVE-…5.1???? CVE-…9.1?v2? v3? Excel is the system of record WHAT COMES OUT Same finding, three scores scanner 7.8 · pentest 5.1 · research 9.1 CVSSv4 attempted, not defended no rubric, no rationale captured Bob ≠ Joe different criteria, different answers, one audit Backlog grows every week the CVE curve climbs Designed for a trickle. Fed by a firehose.
Five inbound feeds, one human normalizer, and a spreadsheet in the middle. Every arrow is manual effort that scales linearly while discovery scales exponentially.
The pointA spreadsheet is not a vulnerability management system. It is where triage rationale goes to die, and it is the first thing an auditor asks to see.
Problem Space · 03 · Premarket Submission Risk

Premarket testing is not "let's do a pentest"

FDA's premarket cybersecurity guidance expects security testing to live inside your Secure Product Development Framework (SPDF). That means periodic testing throughout development, not one engagement before submission.

Under the SPDF model, testing is supposed to find and fix vulnerabilities during development. By the time your final reports reach an FDA or MDR submission, they should show little to nothing open, and a lot of history: prior discovery, remediation, and re-verification across builds.

And when no vulnerabilities remain, the report still has to prove you looked. That means an enumerated list of every cybersecurity test case launched against the product's scope, so a reviewer can see coverage of the attack surface even where the finding count is zero. A single late pentest produces the exact opposite: one snapshot, open findings, no fix history, no coverage story.

SPDF · PERIODIC TESTING ACROSS DEVELOPMENT Design Test wave 1 14 found14 fixed Alpha build Test wave 2 31 found29 fixed · 2 mitigated Beta build Test wave 3 12 found12 fixed Release candidate Test wave 4 3 found3 fixed Final regression Full re-run 0 open SUBMISSION PACKAGE FDA 510(k)/PMA · EU MDR 0 open vulnerabilities 60 found and fixed, with history 1,847 test cases enumerated coverage mapped to scope re-verification per build Evidence of process, not a snapshot THE ANTI-PATTERN One pentest, two weeks before submission One snapshot · findings still open at filing · no fix history · no enumerated coverage · deficiency letter risk
What a reviewer wants to see: findings discovered and burned down across development, zero open at filing, and an enumerated test case list proving coverage where nothing was found.
The pointA clean final report with no testing history reads as untested, not secure. Coverage is the deliverable. Findings are just the byproduct of getting there.
Problem Space · 04 · Premarket Submission Risk

Threat mitigation testing is not V&V

FDA's premarket cybersecurity guidance specifically requires testing of threat mitigations: evidence that the security controls your architecture claims actually work. Checking that a requirement exists is not the same as attacking the control that implements it.

Traditional V&V asks "is the requirement present?" Threat mitigation testing asks a harder question: is the control implemented correctly, and does it hold up against the threat it was designed to stop? Those are different disciplines with different evidence.

The only practical way to do this at device scale is a digital twin that enumerates the security controls where they exist in your documentation, overlays them on the architecture, and employs AI to create a test plan and execute test cases against each control. The output is test case evidence per control: implementation confirmed, efficacy measured, and any associated weakness or vulnerability captured against that control.

YOUR DOCUMENTATION Security architecture views Risk assessment · design specs Claimed controls, extracted AI TEST PLAN · PER CONTROL Controls enumerated and overlaid on the twin Test cases generated against each control executing… DIGITAL TWIN · CONTROLS OVERLAID WHERE THEY LIVE Secure bootbootloader · verify chain AuthN / session mgmtUI + API · roles, lockout TLS everywherecloud + HL7 egress Kiosk lockdowntouchscreen shell escape? Encryption at restPHI store · key handling PASS · evidence attached TC-0412 secure boot chain verified · log + capture WEAK · efficacy partial TC-0688 lockout bypass after 40 attempts FAIL · finding raised TC-0701 kiosk escape to OS shell · vulnerability
Controls extracted from documentation, overlaid on the twin, and attacked one by one. Every control ends with test case evidence: implemented and effective, partially effective, or broken with a finding attached.
Traditional V&V

Checks the requirement exists

"The device shall implement authentication." Test passes because a login screen is present. Nothing was attacked. Efficacy unknown.

Threat mitigation testing

Attacks the implementation

Credential stuffing, session fixation, lockout bypass, token replay against that same login. The control earns its place in the architecture with evidence.

The pointFDA asks for proof the mitigation works, not proof the requirement was written down. If your submission's control evidence is a requirements trace matrix, that section is empty.
Problem Space · 05 · Premarket Submission Risk

Threat model drift: when testing proves the model wrong

Everyone has seen it. The threat model is high level, and then cybersecurity testing finds vulnerabilities in places the model says nothing about. Those two artifacts are supposed to be in sync. When they diverge, the drift is visible to a reviewer.

The fix is resolution. Manufacturers should use AI to consume the threat model they have, overlay it on a digital twin, and expand it to significantly more detail: not just the high-level system, but the intimate behavior of each process inside the product. ELTON does this with proper access from the manufacturer, using not just static scans of code and firmware but real threat actor runtime access to the device.

The result is an enumeration of threats and a cybersecurity test plan 10x anything a human could produce. Is that a problem? Is that too many tests now? For a human, yes. Not for ELTON AI, which executes remotely in your lab or ours. The goal is closing the loop: your FDA-required threat model, your cyber testing, and the vulnerabilities you disposition, moving together in concert, with FDA MDDT rating criteria applied throughout.

YOUR THREAT MODEL · AS DOCUMENTED Device1 box Network1 arrow User1 actor Cloud1 box 8 threats · system level · STRIDE on boxes Testing then finds vulns the model never predicted. Model says one thing. Findings say another. That is drift. AI consumes + expands on twin ELTON EXPANDED MODEL · PROCESS-LEVEL RESOLUTION boot svcipc busdb procapi gw kiosk uiupdaterhl7 svcbt stack 87 threats · per process · static + real runtime access Each threat maps to test cases. 10x what a human would write, executed by AI. ONE CLOSED LOOP · IN CONCERT Threat modelFDA required · expanded on twin Cyber testingtest cases from every threat Dispositionsfindings rated with FDA MDDT criteria
Left: the model you file today. Right: the same model expanded to process-level resolution through the twin. Below: threat model, testing, and dispositions recomputing together instead of drifting apart.
The pointA threat model that testing keeps contradicting is not a threat model, it is a liability with a version number. Sync is achievable when the same AI that expands the model also runs the tests.
Problem Space · 06 · Postmarket Audit Risk

The audit you could fail tomorrow

FDA can stop by tomorrow and ask for the last year of vulnerability data and triage for very specific products and releases. The failure modes are already sitting in your spreadsheet.

If a single vulnerability sits untriaged with a default rating that, evaluated against your cybersecurity patient safety criteria, looks like uncontrolled risk, you have failed. It does not matter that the severity is wrong and was simply never adjusted. You missed it, and the record shows you missed it.

For every finding that arrived as a CVSSv3 HIGH or CRITICAL and was reduced (which is common, and often correct), the reduction needs documented rationale. No justification, and FDA can take issue with the change. And when Bob's rating and Joe's rating differ because they triaged with different criteria, that inconsistency itself is audit risk. This is why digital twins and automated triage with built-in FDA MDDT criteria for CVSSv3 and CVSSv4 are critical: one set of criteria, every finding, rationale attached automatically.

FindingReleaseArrivedTriage statusRating pathRationale on fileAudit outcome
CVE-2025-31842v2.4.1212 days ago Triaged · Not Affected 9.8 → 2.1MDDT rubric, evidence attachedDefensible
CVE-2025-40117v2.4.1147 days ago UNTRIAGED 9.1 (default)NoneUncontrolled risk · FAIL
CVE-2024-50623v2.3.0301 days ago Triaged · reduced 8.6 → 4.2Missing justificationFDA may challenge
CVE-2025-22391v2.4.189 days ago Triaged twice 7.8 → 6.9 (Bob) · 7.8 → 3.4 (Joe)Two different criteriaInconsistent process
CVE-2025-08834v2.2.740 days ago Auto-triaged 7.5 → 3.1MDDT criteria, auto-generatedDefensible
A year of triage the way an investigator reads it. One untriaged default-critical finding is a failed audit, even if the true severity was low all along.
Failure 1

The one you missed

Untriaged, default rating, screens as uncontrolled risk against your patient safety criteria. The correctness of the rating is irrelevant. The silence is the finding.

Failure 2

The reduction you can't defend

HIGH or CRITICAL reduced with no documented rationale. Right answer, wrong record. FDA takes issue with the change, not the number.

Failure 3

Bob versus Joe

Two engineers, two criteria, two ratings for the same finding. Opinion-based triage reads as no process at all.

The pointAudit readiness is not about having good answers. It is about every finding having the same answer machinery: twin context, MDDT criteria for CVSSv3 and CVSSv4, and rationale generated at triage time, not reconstructed the night before an inspection.
Problem Space · 07 · Postmarket Audit Risk

One CVE. Every commercial release. Separately.

FDA guidance requires every release commercialized by a product to be managed separately. Triaging a vulnerability once "for the product" is not enough.

The same vulnerability has to be evaluated within each release, against that release's unique attributes: its own digital twin, its own component versions, its own mitigations, and its own set of co-resident vulnerabilities that could chain. A disposition that is true for v3.0 can be false for v2.1, and an auditor will ask about both.

Manual teams collapse this to one row per CVE because per-release triage multiplies the workload by the size of the installed base. That shortcut is exactly what an investigator is trained to find.

NEW SBOM CVE CVE-2026-13377 OpenSSL · NVD base 9.8 evaluated per release twin RELEASE v2.1 · FIELDED 2023 · 4,100 UNITS Twin: OpenSSL 1.1.1k reachable on exposed HL7 listener Chains with existing local privesc finding AFFECTED · MDDT 8.1 · patch required RELEASE v2.4 · FIELDED 2025 · 9,800 UNITS Twin: same component, listener bound to localhost Mitigation shipped in v2.4 closes the path NOT AFFECTED · VEX issued with evidence RELEASE v3.0 · FIELDED 2026 · 2,300 UNITS Twin: component upgraded, new BLE interface added Not vulnerable to this CVE, new surface under watch CONDITIONAL · monitored against v3.0 twin PER-RELEASE CONTEXT Release-specific digital twin Component version + config Interfaces actually exposed Mitigations in that build Other vulns it could chain with Fielded population + risk criteria One CVE, three twins, three defensible answers. "Product-level" triage collapses these into one wrong answer.
The same CVE lands three different ways across three commercialized releases. Each answer is only defensible because it was computed against that release's own twin.
The pointPer-release management is where manual programs quietly give up, and it is precisely where FDA looks. If your triage table has one row per CVE, you are documenting the shortcut.
ELTON AI · 01 · The Response

All testing, one tool

ELTON pentested thousands of devices over a decade, then encoded that knowledge into a model customers build on. Point-in-time consulting is legacy: expensive, slow, and unable to scale to iterative device development or the postmarket need for low-cost annual monitoring.

ELTON runs manual and automated continuous vulnerability discovery across every source: code, dynamic scanning, fuzzing, reverse engineering, penetration testing, and SBOM CVE proof-of-concept work. Delivery is your choice: the ELTON managed service via subscription testing, or your own team using self-service TestLink™ appliances.

Testing is executed by experienced penetration testers working with an agentic AI physical device testing loop, custom-built over three years, purposely for medical devices. Here's the part most people miss: it is not about the model. It is about the harness. ELTON built an extensive ability for AI to reason about the attack surfaces ELTON identified, to interact with physical interfaces, and to employ loot obtained in one area of the product to iteratively bypass the next. The model provides reasoning. The harness provides a decade of device muscle memory.

DISCOVERY · EVERY SOURCE, ONE PIPELINE Code · SASTsource and build analysis Dynamic scanninglive interface probing Fuzzingprotocol and input mutation Reverse engineeringfirmware and binary Penetration testingexpert-led, AI-assisted SBOM CVE PoCexploit attempts per CVE AGENTIC AI TESTING LOOP The harness, 3+ years in the making Reasons about identified attack surfaces Interacts with physical interfaces Employs loot to iteratively bypass agents · fuzzers · exploits Experienced device pentesters in the loop, directing and reviewing LOOT LOOP · creds, keys, shells from one area unlock the next ELTON MANAGED SERVICE Subscription testing, continuous Our lab or remotely into yours SELF-SERVICE · TESTLINK Appliance on your bench Same harness, your hands Legacy consulting: one snapshot, big invoice. ELTON: the same expertise, running continuously, at subscription cost.
One pipeline for every discovery source, driven by the agentic harness and expert pentesters, delivered as a managed service or on your own bench through TestLink™.
The pointEvery AI pentest vendor has a model. Almost none have a decade of medical device loot-chaining encoded into a physical testing harness. The harness is the moat.
ELTON AI · 02 · The Response

Cybersecurity testing shifts left with TestLink™

Medical device developers and designers have never had access to their own pentesting capability, for any product or release, on their own bench. TestLink™ changes that.

Connect a TestLink™ appliance to the device. Log in to the platform and associate it with a digital twin. Tell the platform which physical connections the appliance has by selecting the twin component that matches each physical connection. Then push the button: the test plan executes, test case results stream in, and vulnerabilities appear as findings, live.

As testing continues, move the appliance across different interfaces (Ethernet to USB to serial) and execute more of the plan, completing the assessment interface by interface. The engineer who wrote the code watches the results arrive at the same bench where the code runs.

ELTON PLATFORM Digital twin association · test plan · results TC-0412 PASS · secure boot TC-0688 FAIL · finding raised TC-0689 RUNNING… OUT-OF-BAND 5G never touches your network YOUR BENCH · DEV LAB TESTLINK black box appliance ETH · USB · SER · BLE test execution local control plane via 5G DEVICE UNDER TEST any product, any release on the developer's bench ETH0 ↔ twin: HL7 interface USB ↔ twin: service port 1 Cable it connect TestLink™ tothe device interfaces 2 Map to twin pick the twin componentfor each physical port 3 Push the button test plan executes,results stream in live 4 Move and repeat shift interfaces, run more,complete the test
TestLink™ on the bench: physical connections mapped to twin components, the plan executed at a push, results streaming to the platform over out-of-band 5G.
The pointShift-left for security has always meant "run a linter earlier." TestLink™ shifts actual device pentesting to the bench where the product is being built, for every release, not the one release a consultant saw.
ELTON AI · 03 · The Response

Your pentesting fleet: cut testing time by 90%

Give every development team a TestLink™ and cybersecurity testing stops being the queue your release waits in. A fleet of appliances on benches you already own turns pentesting into something developers run tonight, not something you schedule next quarter.

The old economics were brutal: book a consultancy, wait for the slot, ship them hardware, wait for the report, dispute the findings, then book again to re-test the fixes. Weeks per cycle, repeated for every release. A TestLink™ fleet inverts that. Developers run self-directed pentesting the moment a build lands, test plans come from the digital twin, and results stream into the platform while the build is still warm. That is where the 90% goes: the waiting, the shipping, the scheduling, the re-engagement, all gone. Products ship faster because testing moves at the speed of the team, not the vendor's calendar.

And here is the part your regulatory team will ask about first: self-directed does not mean self-invented. Every fleet test executes ELTON's methodology, every finding is rated with the FDA MDDT criteria, and every report retains the ELTON brand and the FDA approved vendor credentials reviewers already recognize from 600+ submissions. Your engineers push the button. The evidence still carries the vendor's name.

90%
Testing time cut
Weeks of scheduling become same-day runs
0
Consultant bookings
Self-directed by your developers
100%
Releases covered
Every build, every bench, same plan
1
Evidence standard
ELTON brand + FDA approved vendor credentials
Click the tbox-0001 row to dive into the box
TESTLINK FLEET
NORTHBOUND OPS · FLEET MAP · MOBILITY
FLEET
SVC testlink-gateway · PORT :9090 · mqtt-wss
POLL #6 · 15:21:54Z · NEXT 15:21:51Z
● GW OK● LIVE
15:21:57
2026-07-17 FRI UTC
FLEET LINK
2/3 conn
✖ 1 down
FLEET HEALTH
2/3 ok
▲ 1 degraded
CONFIG SYNC
2/3 sync
▲ 1 drift
MOBILITY
6/7 online
◆ 1 workspace
RTT P50
731 ms
p95 875 · n=6
FLEETDEPLOYMENTS3 BOXES
IDTENANTSTATEHLTHSYNCSCREEN
tbox-0002Acme Corp
tenant-b
● CONN● OK● v9
tbox-0003Northwind Labs
tenant-c
✖ DISC▲ DEGRAD▲ 5→7
UMRROSTER7 UNITS · 2 PAIRED
TBOX-UMR-Lab ONLINE◆ tbox-0001 · tbox-0002
37.77, -122.42 · LTE GOOD · VPN connected · 3.0/10.0 GiB · Verizon Wireless
TBOX-UMR-Field DEGRADED△ UNPAIRED
40.71, -74.01 · LTE FAIR · VPN disconnected · 8.0/10.0 GiB · AT&T Mobility
TBOX-UMR-Chicago ONLINE△ UNPAIRED
41.88, -87.63 · LTE EXCELLENT · VPN connected · 1.1/10.0 GiB · Verizon Wireless
TBOX-UMR-Denver-1 ONLINE◆ tbox-0003
39.74, -104.99 · LTE GOOD · VPN connected · 4.8/10.0 GiB · AT&T Mobility
TBOX-UMR-Seattle OFFLINE△ UNPAIRED
47.61, -122.33 · LTE NO_SIGNAL · VPN disconnected · 5.6/10.0 GiB · T-Mobile US
CLICK A ROW OR A MAP SITE TO DIVE INTO THE BOX · TBOX LAB
CONUSDEPLOYMENT GRIDSITES 6 · UNITS 7 · OFF-MAP 0
TBOX-UMR-Seattle OFFLINE · UNPAIRED TBOX-UMR-Lab ×2 BOXES · WORST ONLINE ×2 SITE ×2 WORST OFFLINE ×2 TBOX-UMR-Dallas ONLINE · UNPAIRED TBOX-UMR-Chicago ONLINE · UNPAIRED TBOX-UMR-Field DEGRADED · UNPAIRED ◇ BOX·UMR PAIR△ UNPAIRED UMR ● ONLINE● DEGRADED● OFFLINE ○ ACTIVE SESSION×N CO-LOCATED
ELTON-branded mock of the TestLink™ fleet console. Left toggle: the ops deck your program manager watches (fleet link, health, config sync, mobility, deployment grid). Right toggle: one TestLink™ opened, tbox-0001, with twin vitals, the operator-to-DUT ops flow, live latency, UMR mobility, and the event log.
The pointA fleet dashboard is not the product. The product is what it removes: the calendar between your developers and their next pentest. Speed without losing the vendor credentials is the whole trade.
ELTON AI · 04 · The Response

Automated verification: real exploitation, not probability

No more reachability analysis tools. No more probability calculations. The structural advantage of medical devices is that every release is identical globally, so exploitation attempted once answers the question everywhere.

ELTON attempts real exploitation on real runtime access, on demand, as soon as an SBOM CVE is recognized. When the platform identifies a new CVE or SAST finding, it creates a test, identifies the test case to run, and if a TestLink™ or a build environment is connected, attempts the exploit.

The result is a test case outcome that records exactly when exploitation was and was not possible, relative to the test case and the access used. That evidence lets the vulnerability status change immediately and automatically to Not Affected in VEX, and removes the issue from all vulnerability chaining analysis. The disposition is not an opinion about reachability. It is the recorded outcome of an attempt.

1 · RECOGNIZED CVE-2026-13377 SBOM match · or SAST finding 2 · TEST CREATED Test case identified from twin context + threat model 3 · EXPLOIT ATTEMPTED Real runtime access on-demand, non-destructive 4 · RESULT · AUTOMATIC VEX: Not Affected exploitation not possible from tested access · evidence attached removed from chaining analysis ✓ EXECUTION SUBSTRATE · WHATEVER IS CONNECTED TestLink™ · live device Build environment The old way: reachability guess · EPSS probability · analyst opinion None of those are evidence. Attempted once. Answered globally. Every release is the same device everywhere.
CVE recognized, test created, exploitation attempted against whatever substrate is connected, and the VEX status flips automatically with the test case result as evidence.
The pointReachability tools estimate. ELTON attempts. When the attempt fails from the tested access, Not Affected is not a claim you defend, it is a result you show.
ELTON AI · 05 · Focus on the 1%

Redefining what a vulnerability is in a medical device

A vulnerability that is not exploitable from a known attack surface is not a vulnerability. It is a weakness. A weakness can carry any rating it likes; it is not exploitable, so it is not an issue requiring action.

ELTON's platform lets a manufacturer take that position on the record, and defend it, because ELTON can defensibly define what is and is not exploitable. The digital twin builds attack paths from known initial access points (the attack surface) to identify what is directly exploitable. Those are the root vulnerabilities, and they are the only findings that are directly exploitable.

Everything gained through a root is an indirect, conditionally exploitable vulnerability: real, but only alive while its enabling conditions are produced upstream. Findings with no path and no met conditions are weaknesses. Three states, computed from the graph, not asserted by an analyst.

THE GRAPH DECIDES · NOT THE ANALYST INITIAL ACCESS Touchscreen UI ROOT · DIRECTLY EXPLOITABLE Kiosk breakout reachable from attack surface DERIV · CONDITIONALLY EXPLOITABLE Local privesc (SBOM CVE) alive only while root produces access DERIV · TERMINAL Credential DB read C-touch C-os-access C-admin WEAKNESS · NOT AFFECTED Firmware signing bypass CVSS 8.2 · needs C-jtag · no producer NO ENTRY VECTOR JTAG not exposed on device condition unmet → no path → no vulnerability, any rating it likes ROOT · directly exploitable · fix these first CONDITIONAL · alive only while the root lives WEAKNESS · no path, no action, evidence attached
One chain, three classifications. Only the root is directly exploitable. The derivative lives on the root's output. The weakness has a critical score and no way to be reached.
The pointThis position only survives an audit if the boundary between vulnerability and weakness is computed, evidenced, and recomputed as the device changes. That is what the twin's attack paths are for. See it live in the next section.
ELTON AI · 06 · Focus on the 1%

FDA recognized ratings, and where severity stops

Every finding is scored first: the FDA MDDT rubric for CVSSv3 and ELTON's own FDA-aligned MDDT methodology for CVSSv4. That is where the land of severity ends. Severity never decides what requires action. Exploitability does.

Who says a CVSS 10 matters when it is not even exploitable? The FDA does, if you are calling it a vulnerability. But it is not a vulnerability if it is not exploitable. That is a weakness. ELTON determines what is directly exploitable based on conditions that are met (keys that are known, credentials, runtime access) and the graph path of the digital twin: which root vulnerability meets conditions that enable the other, indirect vulnerabilities, and which findings have no conditions or paths at all.

Try it below. The kiosk breakout is a root vulnerability. It produces one condition: OS access. Every vulnerability living on the OS needs that condition to be exploitable. Fix the kiosk root and watch the OS findings become weaknesses on the fly. Their isolation scores never change. Their status does.

Toggle the root fixed and the graph recomputes every status
4
Vulnerabilities · affected
1
Weaknesses · not_affected
0
Roots fixed
5
Total findings
INITIAL ACCESS Touchscreen UI ROOT · VULN K1 · Kiosk breakout isolation 7.6 High · MDDT rated DERIV · VULN K2 · Local privesc isolation 7.8 High · SBOM CVE DERIV · VULN K3 · Read credential DB isolation 6.5 Med DERIV · VULN K4 · Weak KDF in DB isolation 9.1 Critical WEAKNESS · ALWAYS K5 · Firmware signing bypass isolation 8.2 High · C-jtag unmet no producer for C-jtag on this device C-touch C-os-access C-admin C-dbhash The kiosk root is live. Its condition C-os-access is met, so every OS finding downstream is an exploitable vulnerability.
IDFindingRequires → ProducesIsolation CVSS · MDDTRoleStatus
Isolation ratings never move; they stay intact for compliance. Only status recomputes. One fix retires four findings' worth of action, and the rationale for each flip is the graph itself.
The pointSeverity is where triage starts, not where it ends. Rate everything with the MDDT rubrics, then let exploitability, computed from conditions and paths on the twin, decide what is a vulnerability at all. That is how you find the 1% that require fixing and prove the other 99% do not need to be.