Two lab technicians in safety glasses testing medical-device equipment, one at a computer and one adjusting an instrument rack

Medical & regulated devices

DICOM and PACS cybersecurity: why medical imaging still runs unauthenticated by default

Last updated:

DICOM (Digital Imaging and Communications in Medicine), the NEMA-maintained standard behind how every PACS archive, imaging modality and diagnostic workstation exchanges studies, treats authentication, encryption and integrity checking as optional under its own PS3.15 profile, not mandatory by default. A 2019 audit found 590 internet-exposed PACS servers requiring no authentication at all, and CISA's 2026 advisory against the OFFIS DCMTK toolkit shows the same class of gap continuing to surface in current imaging software, not just legacy PACS deployments.

Key takeaways

  • DICOM's base network protocol — the association negotiation and storage/query services (C-STORE, C-FIND, C-GET) that move imaging studies between Application Entities — has no built-in authentication of its own.
  • Two details matter.
  • PS3.15 Annex E defines how a "de-identifier" strips identity before studies go to research, teaching files or registries.
  • The clearest evidence that this optionality translates into real exposure came from a 2019 investigation by security firm Greenbone Networks, working with ProPublica and German broadcaster Bayerischer Rundfunk.

What does PS3.15 actually require, and is any of it mandatory?

DICOM's base network protocol — the association negotiation and storage/query services (C-STORE, C-FIND, C-GET) that move imaging studies between Application Entities — has no built-in authentication of its own. Security is addressed in a separate part of the standard, PS3.15 ("Security and System Management Profiles"), which NEMA's own text frames around the premise that "Application Entities may not trust the communications channel by which they communicate with other Application Entities". But PS3.15 is a menu, not a requirement: the standard states an implementation "may claim conformance to any of the Security and System Management Profiles individually," and nothing in base DICOM forces a PACS or modality to implement any of them. The Online Electronic Storage Secure Use Profile is the one that actually closes the authentication gap for image transfer — mutual authentication over a secure transport connection, with the receiving side required to reject and discard any instance that fails a post-transfer integrity check — but it only applies where a site has deliberately turned it on.

The port numbers show the same split. IANA's service-name registry lists TCP 104 as "acr-nema" and 11112 as "dicom" for plain associations, and a separate port, 2762, as "dicom-tls". A product that only listens on 104 or 11112 is, by construction, running DICOM without transport security.

Which PS3.15 profiles should an imaging product implement?

The short answer for a modality, archive or viewer shipping in 2026: one of the current TLS profiles, an authenticated user identity on the association, the audit trail format with a secure transport for it, and the de-identification profile wherever studies leave the clinical network. The current edition of the standard (PS3.15 2026d) lists these as separate annexes, and several older options are retired:

Need PS3.15 profile Status in 2026d What it requires
Encrypted, authenticated transport B.12 BCP 195 RFC 8996, 9325 TLS Current Comply with BCP 195 (RFC 8996, RFC 9325); servers and clients "shall support TLS 1.2 and may support TLS 1.3"; clients try TLS 1.3 first
Stricter transport for new designs B.13 Modified BCP 195 RFC 8996, 9325 TLS Current Same base plus a permitted-algorithm list: ECDHE/DHE key exchange, AES or Camellia in GCM/CCM, ChaCha20-Poly1305, SHA-256/384; MD5, SHA-1, RC4, 3DES and CBC mode excluded
Older TLS profiles B.1 Basic TLS, B.2 ISCL, B.3 AES TLS Retired (see PS3.3-2018a) Do not claim these for new products
Earlier BCP 195 profiles B.9, B.10, B.11 Retired (see PS3.3-2022d) Migrate claims to B.12 or B.13
User, not only machine, identity B.4 Basic User Identity, B.5 User Identity plus Passcode, B.6 Kerberos, B.7 SAML Current B.4 "need not verify the passcode"; under B.5, "If the authentication fails, the association shall be rejected"
Audit trail A.5 Audit Trail Message Format, A.6 SYSLOG-TLS Current XML audit messages sent over syslog per RFC 5425; "TLS MUST be used"; receivers handle messages of at least 32,768 octets
De-identification E.2 Basic Application Level Confidentiality plus options Current See the next section

Two details matter. First, B.4 accepts a username without checking it, so it is useful for audit records but is not access control; B.5, B.6 or B.7 is what lets an archive reject an association from an unknown user. Second, the TLS profiles describe what a product can do, not how a site configures it: the standard notes that "DICOM does not specify how such devices are configured in the field". The conformance statement and the customer security documentation have to say which profile is on by default.

How does DICOM de-identification work, and where does it fall short?

PS3.15 Annex E defines how a "de-identifier" strips identity before studies go to research, teaching files or registries. The Basic Application Level Confidentiality Profile "defines an extremely conservative approach" that removes patient identity and demographics, staff and organisation identity, UIDs, dates and times, and private attributes. Each attribute gets an action code from Table E.1-1a: D (replace with a dummy value), Z (zero-length or dummy), X (remove), K (keep), C (clean) and U (replace with a consistent UID).

The gap is the pixels. The profile itself says: "Unless the Clean Pixel Data Option or the Clean Recognizable Visual Features Option is specified, this Profile does not address information in the pixels." The Clean Pixel Data Option notes that CT images do not normally carry burned-in annotation, "whereas Ultrasound images routinely do", and that removal "may require intervention of or approval by a human operator". Studies that need dates for longitudinal analysis can use one of the two Retain Longitudinal Temporal Information options, which trade some privacy for usable timelines. A product that exports "anonymised" studies should name the exact profile and options it applies.

Did a real-world audit find this gap at scale?

The clearest evidence that this optionality translates into real exposure came from a 2019 investigation by security firm Greenbone Networks, working with ProPublica and German broadcaster Bayerischer Rundfunk. Between July and September 2019, Greenbone scanned roughly 2,300 internet-connected PACS servers located via Shodan and Censys, and found 590 required no authentication whatsoever to retrieve stored images. Those 590 servers alone exposed an estimated 24.3 million patient studies and 733 million medical images — about 399.5 million directly viewable or downloadable — across 52 countries, with the United States accounting for the largest single concentration (187 servers, roughly 303 million images, about 5 million patients), per Greenbone's report. A November follow-up found exposure had grown to an estimated 35 million patients and 1.19 billion scans. Greenbone's own conclusion was that the root cause was misconfiguration — PACS deliberately exposed so remote physicians could reach them — not a defect unique to any vendor or to DICOM itself, which fits PS3.15's design: running without the secure-use profiles is the path of least resistance.

What do CISA's imaging advisories show about current software?

CISA's ICSMA-26-181-01 (published 30 June 2026, CSAF file) assigned five CVEs to OFFIS DCMTK, an open-source DICOM toolkit embedded in many commercial imaging products (affected versions 3.7.0 and earlier), reported by researcher Abhinav Agarwal:

CVE CWE CVSS v3.1 What it allows
CVE-2026-50003 CWE-22 Path Traversal 9.8 (Critical) A malicious or compromised server can force a DCMTK client in bit-preserving C-GET mode to write files outside its intended output directory
CVE-2026-52868 CWE-22 Path Traversal 8.2 (High) An unauthenticated attacker can read worklist records outside the intended per-AE storage area, crossing departmental data separation
CVE-2026-50254 CWE-401 Memory Leak 7.5 (High) Repeated crafted connection requests exhaust memory on storescp, eventually killing the service until an operator restarts it
CVE-2026-35505 CWE-401 Memory Leak 7.5 (High) Same memory-exhaustion pattern via a different crafted-connection path
CVE-2026-44628 CWE-843 Type Confusion 7.5 (High) A single crafted query crashes the worklist server

DCMTK is not an outlier. Of the 55 ICS medical advisories (ICSMA) dated January 2024 to October 2026 in CISA's CSAF repository, 22 name a DICOM viewer, PACS, DICOM server or DICOM library in their title (5 in 2024, 10 in 2025, 7 in 2026 to date; count made 12 November 2026). The same weakness classes keep returning:

Weakness class Example advisory Product and finding Fix named in the advisory
No authentication in the default configuration ICSMA-26-288-02 dcm4chee-arc-light 5.35.1 "accepts DICOM associations from an unregistered calling AE"; C-FIND and C-GET then enumerate studies (CVE-2026-105279, CWE-306, 9.1) Release 5.35.2 or later
No authentication in the default configuration ICSMA-25-037-02 Orthanc before 1.5.8 "does not enable basic authentication by default when remote access is enabled" (CVE-2025-0896, CWE-306, 9.8) Update, or set "AuthenticationEnabled": true
Path traversal from a dataset value ICSMA-26-176-01 pynetdicom's qrscp C-STORE handler uses an attacker-supplied value in a file path (CVE-2026-56445, CWE-22, 9.1) Maintainer did not respond to CISA; check the project page
Malformed-file parsing ICSMA-25-345-01, ICSMA-26-083-01 GDCM out-of-bounds write (CVE-2025-11266, 6.6), then a memory leak in the fixed 3.2.2 (CVE-2026-3650, 7.5); SimpleITK and medInria were listed as affected through bundling 3.2.2 for the first; no maintainer response for the second
Server-side request forgery in a web viewer ICSMA-26-176-02 OHIF Viewers up to 3.12.0: two default data sources fetch an arbitrary URL (CVE-2026-12473, CWE-918, 8.2) 3.12.2
Cleartext credentials ICSMA-25-224-01 Sante PACS Server web portal "sends credential information without encryption" (CVE-2025-54156, CWE-319, 7.4) 4.2.3 or later

Each toolkit advisory reaches every product that bundles an affected version — the pattern SBOM practice exists to make visible (see firmware SBOM for the general case, and fixing path traversal in embedded firmware for the CWE-22 class). Older, more basic failures haven't disappeared either: CISA's ICSMA-18-037-02 (originally posted 6 February 2018, last revised 13 March 2018) found GE Healthcare's Centricity PACS Server, RA1000 workstation and PACS-IW all affected by CWE-287 Improper Authentication (CVSS v3 9.8) tied to default or hard-coded credentials, among them CVE-2012-6693 (default nasro and nasrw passwords on Centricity PACS 4.0 Server) — and as of that advisory's last revision, GE's own patches remained unavailable for three of the affected systems. Removing hard-coded credentials covers that class.

Worked example: how would a modality team triage a toolkit advisory?

Synthetic example, labelled: the product, versions and decisions below are invented to show the method; only the advisory facts come from ICSMA-26-181-01.

An ultrasound maker ships a cart-based scanner whose Linux image includes a DICOM export service built on DCMTK 3.6.8, plus a worklist client. On 30 June 2026 ICSMA-26-181-01 lists DCMTK 3.7.0 and earlier as affected.

  1. Confirm the componentThe release SBOM shows dcmtk 3.6.8 as a package in the image, so the version is in range for all five CVEs.
  2. Map each CVE to a roleThe scanner uses DCMTK as a C-STORE client and a worklist client, not as storescp or a worklist server. CVE-2026-50254, CVE-2026-35505 and CVE-2026-44628 target server roles; the team records why the shipped configuration does not start those services and keeps the evidence (service list, start-up scripts).
  3. Treat the client-side issue as liveCVE-2026-50003 (9.8) is a client issue in bit-preserving C-GET mode; the team checks whether the export service ever uses that mode, and schedules the update either way because a hospital PACS it connects to could be compromised.
  4. Fix and shipThe advisory points to the DCMTK fix in the project's latest GitHub release snapshot; the team rebuilds, regenerates the SBOM and notes the change in the customer security bulletin.
  5. Tell customers what to do meanwhileUntil the update is installed: restrict the scanner's DICOM peers to the hospital PACS AE titles and addresses, and enable the B.12 TLS profile on the export link if the PACS supports it.

What is a workable starting point?

  1. Confirm no DICOM Application Entity is internet-reachable without a VPN or application gateway in front of itGreenbone's audits trace almost every exposed PACS to a deliberate remote-access opening.
  2. Inventory which DICOM toolkit your PACS, viewer or modality interface actually embeds, in the SBOM like any other third-party component.
  3. Turn on a current TLS profile (B.12 or B.13) with mutual authenticationon links where it's supported, rather than assuming network segmentation alone is sufficient — segmentation reduces exposure, it doesn't authenticate the Application Entities talking to each other. Remove any claim of the retired B.1–B.3 or B.9–B.11 profiles from your conformance statement.
  4. Ship authentication on by defaultThe Orthanc and dcm4chee findings above are both default-configuration findings; an archive that accepts any calling AE title is one firewall mistake away from Greenbone's numbers.
  5. Replace default or hard-coded credentials on legacy imaging systems firstThe GE Centricity findings above are nearly a decade old and still unpatched on some affected systems — this is a known, fixable gap, not a novel one.
  6. Send audit messages over SYSLOG-TLS(A.5 format, A.6 transport) so a hospital can see who queried and retrieved studies; see HHS 405(d) practices for how the hospital side uses such logs.
  7. If building or buying imaging software as a medical device, treat the DICOM interface as a distinct attack surface in premarket testing— testing scoped only to the web viewer misses the association and file-parser findings above. See FDA premarket cybersecurity for medical devices for what FDA's 2025 guidance now expects in that testing evidence, and IEC 81001-5-1 for the lifecycle activities behind it.

What are the common pitfalls?

  • Claiming "DICOM TLS" without naming the profile. A buyer cannot tell a retired B.1 claim from a B.13 claim unless the conformance statement names the annex.
  • Treating the AE title as a password. AE titles travel in clear in plain associations and are often printed on configuration sheets; dcm4chee's CVE-2026-105279 shows what happens when an unknown calling AE is accepted.
  • De-identifying headers but not pixels. Ultrasound and secondary-capture images often carry burned-in names; Basic Profile alone does not remove them.
  • Waiting for a maintainer that does not answer. Two advisories above record no maintainer response; the product team still owns the fix or the compensating control.

SoftWSP's device firmware security assessment checks imaging and other embedded devices for outdated bundled libraries as part of a standard scan: see device firmware security.

Where does AI fit into a PACS or imaging firmware scan?

SoftWSP's firmware assessment of a PACS server, modality or workstation image builds a CycloneDX 1.6 SBOM from the packages the extraction finds and matches each component against NVD, CISA KEV, FIRST EPSS and CISA's CSAF ICS advisories (the same feed that carries the ICSMA files above) — that matching is deterministic, not an AI judgment call. A DCMTK, GDCM or dcm4che build installed as a package shows up as a component; a toolkit compiled statically into a vendor binary may not, and the report lists what was and was not identified. AI's role is narrating the result: drafting report language that separates a version match from a confirmed exposure, citing each advisory, and flagging which matched CVEs are KEV-listed or high-EPSS for priority attention. It does not probe a DICOM Application Entity at runtime, check which PS3.15 profiles a link negotiates, or decide whether a server-side CVE applies to a client-only build — that judgement stays with the engineer who reviews every report before it ships.

This guide draws on NEMA's published DICOM PS3.15 standard pages and CISA's published ICS medical advisories (fetched directly), cross-checked against Greenbone Networks' 2019 study and contemporaneous reporting on it where the original PDF's figures are quoted secondhand. This guide is not a substitute for the DICOM standard itself, CISA's own advisories, or a qualified security assessment of a specific PACS deployment.

Put this guide to work

Frequently asked questions

Does DICOM have authentication or encryption built in?

Only as an optional add-on, not a default. DICOM (Digital Imaging and Communications in Medicine) is the NEMA-maintained standard behind how imaging modalities, PACS archives and viewing workstations exchange studies, and its core network protocol — the association and C-STORE/C-FIND/C-GET services that move images between Application Entities — was designed, per NEMA's own PS3.15 part of the standard, on the assumption that "Application Entities may not trust the communications channel by which they communicate with other Application Entities," yet does not itself mandate authentication, integrity checking or encryption on that channel. Security is addressed in a separate part, PS3.15 ("Security and System Management Profiles"), which the standard itself describes as a set of profiles an implementation "may claim conformance to" — entirely optional. In practice, this means a PACS server or modality can accept a C-STORE request from any Application Entity that knows its AE title and port, with no verification of who is actually sending it, unless an administrator has explicitly turned on one of PS3.15's secure-use profiles or added network-level protection like a VPN or firewall rule.

What does PS3.15 actually require when a site does turn security on?

PS3.15 defines several Secure Use Profiles an implementation can conform to, each addressing a different part of the problem. The Online Electronic Storage Secure Use Profile, per the standard's own Annex A.1, requires that a transfer "occur on a Secure Transport Connection," that "the two Application Entities involved in the transfer shall authenticate each other and shall confirm via the authentication that the other supports" the same profile, and that the receiving Application Entity "reject the storage request and discard the received SOP Instance" if post-transfer integrity checks show it was altered in transit — i.e., mutual authentication plus transport encryption plus tamper detection, together. Separately, the Basic Digital Signatures Secure Use Profile addresses data-at-rest integrity: an implementation storing a SOP Instance must guard it against unauthorized tampering and, where possible, validate any digital signature already embedded in it. A third profile defines an Audit Trail Message Format — an XML schema for logging who accessed what imaging data, when, and from where, the kind of record HIPAA-covered entities need regardless of whether DICOM's own encryption is in use. None of this is enabled by default in the base protocol; a site has to choose to implement it, which is exactly the gap that lets a PACS ship and run with no authentication on its DICOM interface at all.

How bad is this in practice — is this a theoretical gap or a documented one?

Documented at a scale that is hard to overstate. In a 2019 audit, security firm Greenbone Networks — working with reporting partners ProPublica and Bayerischer Rundfunk — scanned roughly 2,300 internet-connected PACS servers discovered via Shodan and Censys between July and September 2019, and found 590 of them required no authentication whatsoever to retrieve images. Those 590 servers alone exposed an estimated 24.3 million patient studies and 733 million medical images, of which about 399.5 million could be directly viewed or downloaded, spanning 52 countries, with the single largest concentration (187 servers, roughly 303 million images covering about 5 million patients) found in the United States. A November 2019 Greenbone follow-up found the number of exposed records had grown by more than half, to roughly 35 million patients and 1.19 billion scans. Greenbone's own conclusion was that the root cause was server and network misconfiguration — PACS deliberately exposed so remote physicians could reach them — rather than a flaw unique to DICOM or any single vendor's PACS software, which matches PS3.15's design: the standard offers the secure-use profiles, but nothing in base DICOM stops a site from running without them.

What should a hospital, imaging vendor or PACS integrator actually check first?

Three things, in rough order of what stops the most damage fastest. First, confirm no DICOM Application Entity — PACS archive, modality or viewer — is reachable from the public internet without a VPN or equivalent gateway in front of it; Greenbone's audits found exposure almost always traces to a PACS deliberately opened for remote access rather than an exotic attack path, so this is the single highest-value check. Second, patch imaging software and the DICOM toolkits it embeds: CISA's ICSMA-26-181-01 (published 30 June 2026) assigned five CVEs to OFFIS DCMTK — an open-source DICOM toolkit embedded in many commercial imaging products — including CVE-2026-50003 (CWE-22 Path Traversal, CVSS v3.1 9.8) where a malicious or compromised server can make a DCMTK client write files outside its intended output directory, and CVE-2026-52868 (CWE-22, CVSS v3.1 8.2) letting an unauthenticated attacker read worklist records across departmental boundaries — the kind of finding that reaches every product bundling that toolkit, not one vendor's codebase. Third, for anything older still relying on hard-coded or default credentials: CISA's ICSMA-18-037-02 (originally posted 6 February 2018, last revised 13 March 2018) covers GE Healthcare's Centricity PACS Server, RA1000 workstation and PACS-IW, all affected by CWE-287 Improper Authentication findings (CVSS v3 9.8) tied to default or hard-coded credentials, for which GE's own patches remained unavailable for three of the affected systems as of that revision — exactly the class of legacy gap [PSIRT and disclosure practices](/guides/first-psirt-services-framework/) exist to close before an advisory, not after a breach. Where a product is itself a medical device under FDA's 2025 premarket cybersecurity guidance, a PACS or imaging workstation's DICOM interface specifically needs its own authentication and penetration-testing evidence — see [FDA premarket cybersecurity for medical devices](/guides/fda-premarket-cybersecurity-medical-devices/) — rather than being covered implicitly by the product's general web-application security testing.

Sources

  1. NEMA/DICOM Standard PS3.15: Security and System Management Profiles (current edition)
  2. Greenbone Networks: "Unprotected patient data on the internet" study (2019, PDF)
  3. CISA ICS Medical Advisory ICSMA-18-037-02: GE Medical Devices (last revised 13 March 2018; CVE-2012-6693 et al.)
  4. CISA ICS Medical Advisory ICSMA-26-181-01: OFFIS DCMTK Toolkit (30 June 2026; CVE-2026-50003 et al.)
  5. IANA Service Name and Transport Protocol Port Number Registry (acr-nema 104, dicom 11112, dicom-tls 2762)
  6. CISA CSAF repository on GitHub: ICS and ICS medical advisories (csaf_files/OT/white)
  7. CISA CSAF ICSMA-26-288-02: J4Care dcm4che (15 October 2026; CVE-2026-105279 et al.)
  8. CISA CSAF ICSMA-25-037-02: Orthanc Server (6 February 2025; CVE-2025-0896)
  9. CISA CSAF ICSMA-26-176-01: pydicom pynetdicom Library (25 June 2026; CVE-2026-56445)
  10. CISA CSAF ICSMA-26-176-02: OHIF Viewers DICOM (25 June 2026; CVE-2026-12473)
  11. CISA CSAF ICSMA-25-345-01 and ICSMA-26-083-01: Grassroots DICOM (GDCM)
  12. CISA CSAF ICSMA-25-224-01: Santesoft Sante PACS Server (12 August 2025)
Was this guide helpful?

In this cluster

DICOM and PACS cybersecurity: why medical imaging still runs unauthenticated by default