Skip to content
Zero Day Playbook

After patching, hunt the exposure window for what the patch leaves behind

A compromise is a judgment on the evidence

NIST's glossary, quoting CNSSI 4009-2022, defines a compromise as "a judgment, based on the preponderance of the evidence, that a disclosure of information to unauthorized persons or a violation of the security policy for a system in which unauthorized, intentional or unintentional disclosure, modification, destruction, or loss of an object has occurred." A compromise assessment is the work that produces that judgment after a patch: it searches the exposure window and writes what the evidence supports.

The window opens at the advisory, and earlier if you ran the chained bugs

JFrog and NVD published CVE-2026-82329 on 28 Aug 2026. Fastly saw no exploitation that day. It counted only single-digit probes on 29 and 30 Aug, then about 75,000 requests on 31 Aug, nearly 98% from offensive-security services. A public exploit appeared on 1 Sep, with about 171,000 attempts. The peak came on 2 Sep, at about 406,000. CISA added the bug to its catalog that day. Wiz confirmed successful exploitation by several actors from 1 to 8 Sep.

Wiz also saw CVE-2026-42018 and CVE-2026-42016 chained on Artifactory from 15 Aug 2026. If an instance ran builds exposed to those two as well, start the window on 15 Aug. The example below assumes the instance was exposed only to CVE-2026-82329 and was upgraded on 3 Sep 2026 (an assumption for the walk-through, not a reported event). That gives a seven-day window, 28 Aug to 3 Sep.

The window does not end on patch day. Fastly states that patching does not revoke tokens already issued, so an admin token minted on 2 Sep still works on 20 Sep. Search for use of tokens and accounts from 28 Aug until the day they are revoked.

Date (2026)Public activity in the vendor and research reportsWhat the date tells your search
28 AugAdvisory and NVD record published; Fastly sees no exploitationEarliest day to search; also the date from which to revoke tokens
29 and 30 AugSingle-digit probes (Fastly)A 201 here would be an early, targeted hit, not noise
31 AugAbout 75,000 requests, nearly 98% from offensive-security servicesExpect many rejected requests; only a 201 matters
1 SepPublic exploit; about 171,000 attempts; Wiz's successful exploitation beginsHighest-value day for new admin users and plugins
2 SepAbout 406,000 attempts; CISA adds the bug to KEVPeak volume; look for the same source repeating across many hosts
3 SepFastly publishes its analysis; the example instance is upgradedCopy evidence before the restart; see the next section
5 SepCISA due date for federal agenciesReference point if your policy follows the catalog
8 SepLast day of Wiz's observed successful exploitationWindow for reported attacks ends; yours runs until tokens are revoked

Copy evidence off the host before the upgrade

The upgrade restarts the service, and attackers may truncate request*.log files (Wiz), so a log copied late can be a log already cut. Before the upgrade, copy these to storage the Artifactory host cannot write to:

  • The access log of the reverse proxy or load balancer in front of Artifactory, for 28 Aug to the copy time. Edge logs survive an attacker who cleans the host.
  • Every request*.log under the Artifactory log directory, with file sizes and modification times noted.
  • A listing of $JFROG_HOME/var/etc/artifactory/plugins/ with timestamps.
  • The system.yaml file, to show whether additionalJoinKeys was set.
  • An export of the user list, including admin flags and any SSH keys attached to users.

Six searches, each with an owner and a record

Run them in this order. Log formats differ, so adjust the field positions to your proxy. The indicator values come from Fastly and Wiz, read 19 Nov 2026.

Search 1, the entry call. Owner: security lead. In the edge log, find POST requests to /access/api/v1/registry/join, /access/api/v1/registry/join/router and the ?override=true form. Example: grep -E 'POST [^ ]*/access/api/v1/registry/join' edge-access.log | grep ' 201 '. The key ID (kid) sits in the header of the token in the POST body, so it shows only where bodies are captured (a WAF, a proxy with body logging, a packet capture). Fastly treats a 201 carrying the kid e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855, the SHA-256 of an empty string, as high-confidence evidence of compromise. A normal cluster peer signs with a real join key and shows a different kid. Wiz saw 200 or 201 with an admin-scoped token. Without body capture, any 200 or 201 on these paths from an address outside your own nodes needs an explanation. Record: the command, the log files and dates, the hit count.

Search 2, source addresses and user agents. Owner: security lead. Compare sources with the sample high-volume addresses Fastly lists: 213.136.72[.]87, 89.117.49[.]147, 45.33.75[.]219, 23.92.20[.]11, 173.255.226[.]148, 195.154.157[.]234, 62.210.87[.]212, 45.79.190[.]169 and 163.172.132[.]245, and with the user agents jfrt/0/7.161.19, JFrog-Access/7.141.17 and JFrogArtifactory/7.146.25. Both are easy to change, so a miss proves little. Record: the list compared and its source.

Search 3, follow-on calls. Owner: security lead. Look in the same logs for the calls Wiz lists: GET /api/system/configuration, PUT /api/security/users/ with an admin flag in customData, token minting at /access/api/v1/tokens, reads of /access/api/v1/system/security/join_key, and enumeration of users, repositories and tokens. Record: each call, time and source address.

Search 4, accounts. Owner: application owner, who knows the real users. Export users and compare with the identity provider, including admin flags and attached SSH keys. Account names Wiz saw: Nxploited_ plus 3 characters, labadmin_ plus hex, 0xTerror, svc_ plus 8 characters, and look-alikes such as jfrog-distribution, jfrog-insight and repo-service. Record: each unexpected user with its creation time and the decision taken.

Search 5, plugins and filters. Owner: platform engineering. Run find "$JFROG_HOME/var/etc/artifactory/plugins" -type f -newermt '28 Aug 2026' -ls and look for the Groovy plugins Wiz names: rce.groovy, cmd.groovy, jwtToken.groovy, uploadFile.groovy, JFrogImage.groovy, metrics.groovy and httpSession.groovy, which are run at /artifactory/api/plugins/execute/. A broken plugin API, or requests to registry/join, api/plugins or systemUpload rejected with HTTP 400, can mean an in-memory filter named traceFilter is installed (Wiz). The upgrade leaves plugins and filters in place; restarting the service from a clean state clears the filter, but a file on disk stays. Record: file names, timestamps and hashes.

Search 6, log integrity. Owner: platform engineering. Run find -name 'request*.log' -size 0 -newermt '28 Aug 2026'. Empty or missing request logs are themselves an indicator (Wiz). Record: file sizes and times against the copies taken before the upgrade.

The finding decides what you rotate and whether the incident plan starts

Rotate the platform join key and revoke access tokens issued since 28 Aug 2026 in every case. That advice comes from Fastly and does not depend on the finding, because a clean search of logs that an attacker could have cut cannot rule out a token. Write the finding as what was searched, over which dates, with which result.

FindingWhat the searches showedAction
No evidence of accessNo 200 or 201 on the join paths from unknown sources; no unknown users, plugins or filters; request logs intactRecord the logs and dates searched. Rotate the join key and revoke tokens since 28 Aug.
Attempts without successJoin requests from unknown sources, all rejectedAs above, and restrict network access to Artifactory and the Access paths to trusted systems (Wiz).
CompromiseA 201 with the empty-key ID, an unknown admin user, an unexplained plugin, or cut logsStart the incident plan built on NIST SP 800-61r3. Rotate credentials stored in Artifactory, review artifacts published or changed since 28 Aug, and, if private.key was read, reset the token certificate. A patch does not revoke forged tokens.

Other products leave different traces

The same method fits any edge product. Replace the six searches with the vendor's own indicators from its record, keep an owner and a record for each, and apply the same finding table.

Where compromise may have reached the host itself, rebuilding is the vendor's route. Cisco strongly recommends re-imaging suspected ISE nodes, and Mandiant's guidance for BeyondTrust is to rebuild from a clean patched image rather than clean in place.

Product and bugExposure window startIndicator named in the record
BeyondTrust Remote Support, CVE-2026-10276429 Aug 2026, 47 days before the advisory of 15 Oct 2026Files mangoapi.php, /tmp/ingredi_ws and /usr/bin/dbus-authd; missing /ns/log/auth.log and /ns/log/http-access.log; user_changed events raising roles to Administrator
Cisco ISE, CVE-2026-7646016 Sep 2026, the advisory day; Cisco gives no earlier dateSuspicious usernames in the ise-kong access.log on every node; unexpected uploads from the node in firewall logs

Who owns each part, and where assessments fail

The security lead owns the finding. Platform engineering exports the logs and snapshots. The application owner says what normal looks like, such as which addresses are real cluster peers. The record is the evidence folder, the six search results with commands and dates, and the written finding, kept with the incident record.

Assessments usually fail in three ways. The logs were copied after the restart. The window was set from the patch date instead of the exposure date. Or a clean result was written as no compromise instead of what was searched. For Artifactory Cloud, JFrog fixed the service itself, so the self-hosted steps above do not apply.

Sources

  1. JFrog security advisories, CVE-2026-82329 section and workaround, read 19 Nov 2026
  2. Fastly Security Research on CVE-2026-82329, 3 Sep 2026, read 19 Nov 2026
  3. Wiz Research on Artifactory exploitation, 10 Sep 2026, updated 17 Sep 2026, read 19 Nov 2026
  4. CISA Known Exploited Vulnerabilities catalog entry for CVE-2026-82329, read 19 Nov 2026
  5. NVD record for CVE-2026-82329, read 19 Nov 2026
  6. BeyondTrust blog on BT26-05, read 19 Nov 2026
  7. Mandiant and GTIG analysis of the BeyondTrust zero-day, read 19 Nov 2026
  8. Cisco Security Advisory cisco-sa-ISE-ABP-VNSW7Tn5, read 19 Nov 2026
  9. NIST SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cyber Risk Management, read 19 Nov 2026
  10. NIST CSRC glossary, compromise (CNSSI 4009-2022), read 19 Nov 2026

Last reviewed 19 Nov 2026.