GCVE-1988-2026-0502

Vulnerability from gna-1988 – Published: 2026-11-02 07:30 – Updated: 2026-11-02 07:30
VLAI
Title
[0day-rubbish] Moxa MAR-2000 firmware FWR_MAR2000_V1.0_Build_15010909 command injection in the Executer program activation path (7.2)
Summary
0day Rubbish Research Team is publicly disclosing a vulnerability in the Moxa MAR-2000 series industrial 4G/LTE cellular router, firmware image FWR_MAR2000_V1.0_Build_15010909, from Moxa Inc. (Taiwan). An authenticated administrator converts one stored configuration field into arbitrary command execution as the identity of the product's model process, because the Executer model builds the launch command of an operator program as a shell string whose program-argument column is never validated by any layer of the backend. This is one of three independent command injection findings documented for this product and released together as a set. Type and root cause. CWE-78, OS command injection, with CWE-20 co-equal, in the Executer model at /home/model/system/executer/executer.py inside the ARM Debian 6.0.8 root filesystem of the firmware, which runs the vendor's MIRF (Mobile Intelligent Routing Framework) stack: a Node.js Express 4.0 management application on port 80 bound to all interfaces, the express-sanji middleware bridging REST resources onto a local MQTT broker, and Python 2.6 model scripts that persist configuration and act on it. The model stores one row per operator program in an SQLite table whose columns include fileName and argument. On activation it assembles the command as str(execute_file) + " " + str(argument) and launches it with subprocess.Popen(full_command, shell=True, preexec_fn=os.setsid), where the row's argument bytes are concatenated verbatim and shell=True delegates the parsing to /bin/sh: semicolons, background operators, pipes, backticks and command substitution all have their ordinary meaning. The value's provenance is a database column set by an HTTP request: the row writer is generic, taking every key of the PUT /api/v1/system/executer/program/<id> request's data object and committing it with a plain SQL UPDATE ... SET ... statement, with no per-column semantics anywhere on the write path. The only gate on execution is a liveness check - the file must exist under the user-program directory - and the product's own multipart upload route, POST /api/v1/system/executer/program, satisfies it: the upload applies only a filename normalisation, with no allowed-type list and no file filter. The chain is therefore: log in, upload a benign script through the product's own route, store "; <command>; #" in the row's argument field, and activate the row with PUT /api/v1/system/executer/activatedProgram carrying programId and runBit 1. The shell runs the uploaded script first, the semicolon ends it, the injected statement runs, and the trailing hash comments out the rest of the line. The same model contains a second, structurally identical sink in its deletion path: it builds "rm -rf %s" from the row's unvalidated fileName column and executes it with subprocess.call(..., shell=True), guarded only by a comparison of the raw value against a single slash, which blocks exactly one value and nothing else, and requires no upload and no activation. The two-request chain looks like this after login: ``` [1] POST /api/v1/system/executer/program multipart, benign script -> stored under the user-program directory [2] PUT /api/v1/system/executer/program/<id> {"data": {"fileName": "<uploaded basename>", "argument": "; <command>; #"}} -> generic row writer commits both columns verbatim (SQLite) [3] PUT /api/v1/system/executer/activatedProgram {"data": {"programId": <id>, "runBit": 1}} -> Popen("<dir>/<file> ; <command>; #", shell=True) ``` Authentication and reach. The validateUser middleware is the only authentication check on the REST surface: it requires a server-side session for every non-GET method, and login is POST /api/login compared against bcrypt-hashed entries in the SQLite user table. Past that check there is no second authorisation layer - the message bridge performs no permission translation and the broker leg carries no identity - so the product's single documented role, the device administrator, is the whole security boundary. The listener binds all interfaces on port 80 in the shipped firmware; no WAN exposure is assumed anywhere in the published rating. Severity. The published score is 7.2 High, CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:U/C:H/I:H/A:H, from ISS = 0.914816, impact 5.873119, exploitability 1.234708 and base 7.107826 rounded up to 7.2, Privileges Required High being the product's single administrator role. A labelled conditional and never the headline: where a deployment still runs the factory default administrator pair carried by the shipped firmware's SQLite user table, the prerequisite collapses and the chain reads 9.8 Critical on the PR:N vector (exploitability 3.887043, base 9.760161); a changed credential removes it, and the default pair is disclosed as a separate weakness, never folded into the published vector. A PR:L reading of 8.8 High is published only as the arithmetic of the rejected variant. The score provenance is disclosed in public: the research record behind this advisory carried 8.8 for all three same-product findings, that figure is the arithmetic of a PR:L vector, no artefact shows a lower-privileged product role, and every published number was recomputed from its own vector - which lowered the researchers' own figure by 1.6 points. A Scope-Changed reading computes to 9.1 and is deliberately not published, because the model process belongs to the same single-purpose appliance authority, and commanding the appliance's own operating system through its management plane is the product's design rather than an escape. Verification method, and its boundary. Two evidence classes carry the finding, kept separate in the published advisory. The firmware's own application sources establish the entry path, the storage semantics and the exact sink construction, read from the shipped files. The dynamic layer is a sink-level replication on an isolated cloud Linux sandbox: the model's exact concatenation with the shell=True launch, carrying a benign upload stand-in, an injection payload and a benign control; with the payload in the argument position a marker file containing uid=0(root) was created, and the control produced the benign program launch and nothing else. The deletion construction was replicated with the same method and also produced its marker. The boundary is stated plainly: end-to-end detonation against powered-on hardware is outside this research record, and the identity of the model process on the device was neither measured nor claimed - uid=0(root) is the replication sandbox's identity and nothing broader rests on it. Impact. Arbitrary command execution as the identity of the Python model process, begun by one authenticated configuration write and one activation request, with no memory corruption and no protocol manipulation in the chain. The command line is a stored configuration value that persists across reboots and re-executes on every activation of the row, so the payload is reboot-durable without persistence work by the attacker. The deletion sink extends the class to a field reachable without any upload and allows destructive filesystem operations wherever the injected metacharacters lead. The compromise yields the router's complete configuration, control of its forwarding and cellular function, and availability at will over a device deployed as communication infrastructure. The two requests that matter are indistinguishable in shape from ordinary administrator activity, so the readable signals are content-level: program rows whose argument contains shell metacharacters, unexpected files in the user-program directory, unexpected child processes under the model daemon. The web stack's separate property of answering unauthenticated GET requests across the API surface is a read-exposure issue reported on its own and is not part of this chain. Remediation: execute activated programs without a shell, with the executable path and the argument as separate execv-style arguments; validate the argument at the row-write boundary in the model rather than only in a web form; constrain fileName to the set of stored uploads and associate rows to files by an opaque server-side handle; rebuild the deletion path with an argument vector; restrict the multipart upload to executable-basename types and add a regression test that writes a metacharacter-bearing argument, activates the row, and asserts rejection at write time or execution without shell interpretation. Operators until a patch exists: change the factory default administrator credential at commissioning, keep the management listener off WAN-facing addresses, audit the program table and user-program directory on a schedule, and disable the Executer feature where operator programs are not used. Proof of concept. The published tool is a standard-library-only Python 3 client that drives the full chain: it logs in, uploads a benign two-line shell script through the multipart route, writes the argument payload into a chosen row, activates the row, and reports; it verifies both the write and the activation responses and exits non-zero with an explicit message at each failure point. Target, credential, row id, marker command and timeouts are arguments; the default target is loopback. No persistence, no reverse shell, no destructive action; the default command writes a single marker file so the outcome is observable without a second channel. Use it only against systems you own or are authorised to test. The vendor's product security team mailbox, published on the vendor's own security advisories page, was used for the unconditional notification; no report form was submitted and no individual was addressed by name. This disclosure is unconditional: the research is published, no permission was asked for, no embargo was requested, nothing is withheld pending a reply, and remediation questions are answered at no cost. No vulnerability identifier has been assigned to this finding. Two companion advisories for the same product are published alongside this one, independent findings that share only the entry architecture - a different model file, REST resource, configuration field and sink each: an IPsec tunnel address injection through the route-management shell commands, and a scenario ping condition injection through the auto-executing ping thread, both at the second and third URLs below. Full technical analysis and the reproducible proof of concept, published unconditionally: https://0day-rubbish.com/blog/mar-2000-executer-program-argument-command-injection Companion advisories for the same product, independent findings: https://0day-rubbish.com/blog/mar-2000-network-ipsec-command-injection https://0day-rubbish.com/blog/mar-2000-scenario-ping-command-injection Project archive (ongoing disclosure series): https://github.com/Exploit-Garbage/0day-Rubbish -- 0day Rubbish Research Team disclosure () 0day-rubbish com https://0day-rubbish.com _______________________________________________ Sent through the Full Disclosure mailing list https://nmap.org/mailman/listinfo/fulldisclosure Web Archives & RSS: https://seclists.org/fulldisclosure/
Severity
No CVSS data available.
Assigner
VULNARCHIVE GNA GNA-1988
GNA scorecard E 38/100 over 581 records in the last 180 days details
Impacted products
Vendor Product Version CPE status
Moxa MAR-2000 firmware Affected: unknown
guessed Create a notification for this product.

{
  "containers": {
    "cna": {
      "affected": [
        {
          "product": "MAR-2000 firmware",
          "vendor": "Moxa",
          "versions": [
            {
              "status": "affected",
              "version": "unknown"
            }
          ]
        }
      ],
      "credits": [
        {
          "lang": "en",
          "type": "finder",
          "value": "disclosure via Fulldisclosure"
        }
      ],
      "descriptions": [
        {
          "lang": "en",
          "value": "0day Rubbish Research Team is publicly disclosing a vulnerability in the Moxa\nMAR-2000 series industrial 4G/LTE cellular router, firmware image\nFWR_MAR2000_V1.0_Build_15010909, from Moxa Inc. (Taiwan). An authenticated\nadministrator converts one stored configuration field into arbitrary command\nexecution as the identity of the product\u0027s model process, because the\nExecuter model builds the launch command of an operator program as a shell\nstring whose program-argument column is never validated by any layer of the\nbackend. This is one of three independent command injection findings\ndocumented for this product and released together as a set.\n\nType and root cause. CWE-78, OS command injection, with CWE-20 co-equal, in\nthe Executer model at /home/model/system/executer/executer.py inside the\nARM Debian 6.0.8 root filesystem of the firmware, which runs the vendor\u0027s\nMIRF (Mobile Intelligent Routing Framework) stack: a Node.js Express 4.0\nmanagement application on port 80 bound to all interfaces, the express-sanji\nmiddleware bridging REST resources onto a local MQTT broker, and Python 2.6\nmodel scripts that persist configuration and act on it. The model stores\none row per operator program in an SQLite table whose columns include\nfileName and argument. On activation it assembles the command as\nstr(execute_file) + \" \" + str(argument) and launches it with\nsubprocess.Popen(full_command, shell=True, preexec_fn=os.setsid), where the\nrow\u0027s argument bytes are concatenated verbatim and shell=True delegates the\nparsing to /bin/sh: semicolons, background operators, pipes, backticks and\ncommand substitution all have their ordinary meaning. The value\u0027s\nprovenance is a database column set by an HTTP request: the row writer is\ngeneric, taking every key of the PUT /api/v1/system/executer/program/\u003cid\u003e\nrequest\u0027s data object and committing it with a plain SQL\nUPDATE ... SET ... statement, with no per-column semantics anywhere on the\nwrite path. The only gate on execution is a liveness check - the file must\nexist under the user-program directory - and the product\u0027s own multipart\nupload route, POST /api/v1/system/executer/program, satisfies it: the upload\napplies only a filename normalisation, with no allowed-type list and no file\nfilter. The chain is therefore: log in, upload a benign script through the\nproduct\u0027s own route, store \"; \u003ccommand\u003e; #\" in the row\u0027s argument field, and\nactivate the row with PUT /api/v1/system/executer/activatedProgram carrying\nprogramId and runBit 1. The shell runs the uploaded script first, the\nsemicolon ends it, the injected statement runs, and the trailing hash\ncomments out the rest of the line. The same model contains a second,\nstructurally identical sink in its deletion path: it builds\n\"rm -rf %s\" from the row\u0027s unvalidated fileName column and executes it with\nsubprocess.call(..., shell=True), guarded only by a comparison of the raw\nvalue against a single slash, which blocks exactly one value and nothing\nelse, and requires no upload and no activation.\n\nThe two-request chain looks like this after login:\n\n```\n[1] POST /api/v1/system/executer/program      multipart, benign script\n    -\u003e stored under the user-program directory\n[2] PUT  /api/v1/system/executer/program/\u003cid\u003e\n    {\"data\": {\"fileName\": \"\u003cuploaded basename\u003e\",\n              \"argument\": \"; \u003ccommand\u003e; #\"}}\n    -\u003e generic row writer commits both columns verbatim (SQLite)\n[3] PUT  /api/v1/system/executer/activatedProgram\n    {\"data\": {\"programId\": \u003cid\u003e, \"runBit\": 1}}\n    -\u003e Popen(\"\u003cdir\u003e/\u003cfile\u003e ; \u003ccommand\u003e; #\", shell=True)\n```\n\nAuthentication and reach. The validateUser middleware is the only\nauthentication check on the REST surface: it requires a server-side session\nfor every non-GET method, and login is POST /api/login compared against\nbcrypt-hashed entries in the SQLite user table. Past that check there is no\nsecond authorisation layer - the message bridge performs no permission\ntranslation and the broker leg carries no identity - so the product\u0027s single\ndocumented role, the device administrator, is the whole security boundary.\nThe listener binds all interfaces on port 80 in the shipped firmware; no\nWAN exposure is assumed anywhere in the published rating.\n\nSeverity. The published score is 7.2 High,\nCVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:U/C:H/I:H/A:H, from ISS = 0.914816, impact\n5.873119, exploitability 1.234708 and base 7.107826 rounded up to 7.2,\nPrivileges Required High being the product\u0027s single administrator role. A\nlabelled conditional and never the headline: where a deployment still runs\nthe factory default administrator pair carried by the shipped firmware\u0027s\nSQLite user table, the prerequisite collapses and the chain reads 9.8\nCritical on the PR:N vector (exploitability 3.887043, base 9.760161); a\nchanged credential removes it, and the default pair is disclosed as a\nseparate weakness, never folded into the published vector. A PR:L reading\nof 8.8 High is published only as the arithmetic of the rejected variant. The\nscore provenance is disclosed in public: the research record behind this\nadvisory carried 8.8 for all three same-product findings, that figure is the\narithmetic of a PR:L vector, no artefact shows a lower-privileged product\nrole, and every published number was recomputed from its own vector - which\nlowered the researchers\u0027 own figure by 1.6 points. A Scope-Changed reading\ncomputes to 9.1 and is deliberately not published, because the model\nprocess belongs to the same single-purpose appliance authority, and\ncommanding the appliance\u0027s own operating system through its management\nplane is the product\u0027s design rather than an escape.\n\nVerification method, and its boundary. Two evidence classes carry the\nfinding, kept separate in the published advisory. The firmware\u0027s own\napplication sources establish the entry path, the storage semantics and the\nexact sink construction, read from the shipped files. The dynamic layer is\na sink-level replication on an isolated cloud Linux sandbox: the model\u0027s\nexact concatenation with the shell=True launch, carrying a benign upload\nstand-in, an injection payload and a benign control; with the payload in the\nargument position a marker file containing uid=0(root) was created, and the\ncontrol produced the benign program launch and nothing else. The deletion\nconstruction was replicated with the same method and also produced its\nmarker. The boundary is stated plainly: end-to-end detonation against\npowered-on hardware is outside this research record, and the identity of the\nmodel process on the device was neither measured nor claimed - uid=0(root)\nis the replication sandbox\u0027s identity and nothing broader rests on it.\n\nImpact. Arbitrary command execution as the identity of the Python model\nprocess, begun by one authenticated configuration write and one activation\nrequest, with no memory corruption and no protocol manipulation in the\nchain. The command line is a stored configuration value that persists across\nreboots and re-executes on every activation of the row, so the payload is\nreboot-durable without persistence work by the attacker. The deletion sink\nextends the class to a field reachable without any upload and allows\ndestructive filesystem operations wherever the injected metacharacters\nlead. The compromise yields the router\u0027s complete configuration, control of\nits forwarding and cellular function, and availability at will over a\ndevice deployed as communication infrastructure. The two requests that\nmatter are indistinguishable in shape from ordinary administrator activity,\nso the readable signals are content-level: program rows whose argument\ncontains shell metacharacters, unexpected files in the user-program\ndirectory, unexpected child processes under the model daemon. The web\nstack\u0027s separate property of answering unauthenticated GET requests across\nthe API surface is a read-exposure issue reported on its own and is not\npart of this chain.\n\nRemediation: execute activated programs without a shell, with the\nexecutable path and the argument as separate execv-style arguments; validate\nthe argument at the row-write boundary in the model rather than only in a\nweb form; constrain fileName to the set of stored uploads and associate rows\nto files by an opaque server-side handle; rebuild the deletion path with an\nargument vector; restrict the multipart upload to executable-basename types\nand add a regression test that writes a metacharacter-bearing argument,\nactivates the row, and asserts rejection at write time or execution without\nshell interpretation. Operators until a patch exists: change the factory\ndefault administrator credential at commissioning, keep the management\nlistener off WAN-facing addresses, audit the program table and user-program\ndirectory on a schedule, and disable the Executer feature where operator\nprograms are not used.\n\nProof of concept. The published tool is a standard-library-only Python 3\nclient that drives the full chain: it logs in, uploads a benign two-line\nshell script through the multipart route, writes the argument payload into a\nchosen row, activates the row, and reports; it verifies both the write and\nthe activation responses and exits non-zero with an explicit message at each\nfailure point. Target, credential, row id, marker command and timeouts are\narguments; the default target is loopback. No persistence, no reverse shell,\nno destructive action; the default command writes a single marker file so\nthe outcome is observable without a second channel. Use it only against\nsystems you own or are authorised to test.\n\nThe vendor\u0027s product security team mailbox, published on the vendor\u0027s own\nsecurity advisories page, was used for the unconditional notification; no\nreport form was submitted and no individual was addressed by name. This\ndisclosure is unconditional: the research is published, no permission was\nasked for, no embargo was requested, nothing is withheld pending a reply,\nand remediation questions are answered at no cost. No vulnerability\nidentifier has been assigned to this finding.\n\nTwo companion advisories for the same product are published alongside this\none, independent findings that share only the entry architecture - a\ndifferent model file, REST resource, configuration field and sink each: an\nIPsec tunnel address injection through the route-management shell commands,\nand a scenario ping condition injection through the auto-executing ping\nthread, both at the second and third URLs below.\n\nFull technical analysis and the reproducible proof of concept, published\nunconditionally:\n  https://0day-rubbish.com/blog/mar-2000-executer-program-argument-command-injection\n\nCompanion advisories for the same product, independent findings:\n  https://0day-rubbish.com/blog/mar-2000-network-ipsec-command-injection\n  https://0day-rubbish.com/blog/mar-2000-scenario-ping-command-injection\n\nProject archive (ongoing disclosure series):\n  https://github.com/Exploit-Garbage/0day-Rubbish\n\n--\n0day Rubbish Research Team\ndisclosure () 0day-rubbish com\nhttps://0day-rubbish.com\n_______________________________________________\nSent through the Full Disclosure mailing list\nhttps://nmap.org/mailman/listinfo/fulldisclosure\nWeb Archives \u0026 RSS: https://seclists.org/fulldisclosure/"
        }
      ],
      "problemTypes": [
        {
          "descriptions": [
            {
              "cweId": "CWE-20",
              "description": "CWE-20",
              "lang": "en",
              "type": "CWE"
            },
            {
              "cweId": "CWE-78",
              "description": "CWE-78",
              "lang": "en",
              "type": "CWE"
            }
          ]
        }
      ],
      "providerMetadata": {
        "dateUpdated": "2026-11-02T07:30:04Z",
        "orgId": "4e2abfbf-4a2a-4b76-a4e0-d77c18ba156c",
        "shortName": "VULNARCHIVE"
      },
      "references": [
        {
          "tags": [
            "technical-description",
            "exploit"
          ],
          "url": "https://vuln.freearchive.org/archive/full-disclosure/2026/Oct/52"
        },
        {
          "tags": [
            "technical-description"
          ],
          "url": "https://seclists.org/fulldisclosure/2026/Oct/52"
        },
        {
          "url": "https://0day-rubbish.com"
        },
        {
          "url": "https://0day-rubbish.com/blog/mar-2000-executer-program-argument-command-injection"
        },
        {
          "url": "https://0day-rubbish.com/blog/mar-2000-network-ipsec-command-injection"
        },
        {
          "url": "https://0day-rubbish.com/blog/mar-2000-scenario-ping-command-injection"
        },
        {
          "url": "https://github.com/Exploit-Garbage/0day-Rubbish"
        },
        {
          "url": "https://nmap.org/mailman/listinfo/fulldisclosure"
        },
        {
          "url": "https://seclists.org/fulldisclosure/"
        }
      ],
      "source": {
        "defect": [
          "https://seclists.org/fulldisclosure/2026/Oct/52"
        ],
        "discovery": "EXTERNAL"
      },
      "title": "[0day-rubbish] Moxa MAR-2000 firmware FWR_MAR2000_V1.0_Build_15010909 command injection in the Executer program activation path (7.2)",
      "x_gcve": [
        {
          "recordType": "advisory",
          "relationships": [],
          "vulnId": "GCVE-1988-2026-0502",
          "x_vulnarchive": {
            "archiveUrl": "https://vuln.freearchive.org/archive/full-disclosure/2026/Oct/52",
            "automated": true,
            "contentSha256": "2411179a70ba593d7df9352ac3f4c930deb26f30d8f08c87c5a00196a05d3527",
            "evidenceScore": 11,
            "messageId": "",
            "originalUrl": "https://seclists.org/fulldisclosure/2026/Oct/52",
            "policy": "vulnarchive-1",
            "sourceFormat": "text/html",
            "sourcePublishedAt": "2026-10-27T16:38:41Z"
          }
        }
      ]
    }
  },
  "cveMetadata": {
    "assignerOrgId": "4e2abfbf-4a2a-4b76-a4e0-d77c18ba156c",
    "assignerShortName": "VULNARCHIVE",
    "datePublished": "2026-11-02T07:30:04Z",
    "dateUpdated": "2026-11-02T07:30:04Z",
    "state": "PUBLISHED",
    "vulnId": "GCVE-1988-2026-0502"
  },
  "dataType": "CVE_RECORD",
  "dataVersion": "5.2"
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.


Loading…

Loading…

Loading…

Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.

Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
Vulnerability timeline

Sightings and recorded dates, grouped by day · UTC

JSON
How durations are calculated

Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.

Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.

Events included in the timeline

Loading timeline…


    Loading…

    Loading…

    Loading…

    Related by attack behaviour

    Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.