Seven of Atlassian's 19 fixed builds for CVE-2026-21589 are on lines that leave support by 30 December
Atlassian's bulletin for CVE-2026-21589 lists 19 fixed builds across eight Data Center products; seven sit on release lines whose support ends between 3 and 30 December 2026. Honeypot attempts were reported on 6 and 7 October, and the CVE record's Crowd fix, 7.1.1, disagrees with the bulletin's 7.1.7.
By Parminder Kumar Sharma · LinkedIn · · 36 min read
Update, 11:10 BST on 17 October: CISA listed the flaw on 15 October, and Previdian now counts 616 attempts from 67 addresses
The updates below, the latest written on 12 October, record that CISA's catalogue did not list CVE-2026-21589. That is no longer true. This update prints no address, path, file name or request.
1. CISA listed it. CISA's catalogue lists the flaw, added on 15 October with a due date of 18 October and the forensic triage flag set to Yes; the file records ransomware use as "Unknown". The catalogue file is now version 2026.10.16 (1,748 entries). The listing came 9 days after Previdian's first observed attempt on 6 October (derived). CISA does not say what evidence it used. The due date falls on a Sunday, three days after the listing.
2. The count. Previdian's page, updated on 17 October, shows 616 exploitation attempts from 67 unique addresses in 21 countries, seen by 5 sensors, first observed on 6 October and last observed on 17 October. On 12 October it showed 380 from 58 addresses in 19 countries, so the attempts count is 62 per cent higher (derived). Attempts are scans and probes recorded by one company's sensors: they are not victims.
3. Atlassian. Its bulletin still says affected Cloud products have been patched and that its investigation has not found evidence of exploitation; it says nothing about Data Center installations, which are the ones the fixed builds in this briefing cover. That is a reading of the bulletin, not a finding that no Data Center customer was affected.
4. What this changes. Nothing about the fixed builds or the support dates in the tables below. What changes is the clock: a catalogue listing puts the flaw into the NCSC's action table for catalogued flaws, subject to the questions that table asks, and moves a self-hosted Data Center instance on an unfixed build from "exposed in principle" to "on a three-day federal list". The ordered actions below stand, with the same hunt window.
What changed since the update of 12 October. Sources: CISA's catalogue file (version 2026.10.16), Previdian's page (updated 17 October) and Atlassian's bulletin, read on 17 October 2026
| Item | Read on 12 October | Read on 17 October |
|---|---|---|
| CISA's catalogue | Not listed (version 2026.10.08) | Listed 15 October, due 18 October, forensic triage required (version 2026.10.16) |
| Previdian's count | 380 attempts, 58 addresses, 19 countries, 5 sensors | 616 attempts, 67 addresses, 21 countries, 5 sensors; first seen 6 October, last seen 17 October |
| Atlassian on exploitation | Cloud: no evidence found; Data Center not addressed | Unchanged |
| Victims, post-compromise activity | None reported | None reported in the sources read |
Update, 20:20 BST on 12 October: Previdian's count is 380 attempts from 58 addresses in 19 countries and fading, and watchTowr says most were fingerprinting
This briefing was last updated on 9 October. I have re-read Previdian's page for the flaw, The Hacker News, Atlassian's bulletin and the CISA catalogue. This update prints no address, path, file name or request.
1. The count and its shape. Previdian's page, updated on 12 October, now shows 380 exploitation attempts from 58 unique addresses in 19 countries, seen by five sensors, against 248, 42 and 14 on the morning of 9 October. Its daily chart reads 6 attempts on 6 October, 107 on 7 October, 108 on 8 October, 86 on 9 October, 54 on 10 October, 15 on 11 October and 4 on 12 October, a day that was not over when I read it. The two peak days, 7 and 8 October, hold 215 of the 380 attempts (56.6 per cent, derived), and the daily count has fallen every day since. The first observation is still 6 October and the last is 12 October. It is honeypot traffic: it shows that scanning for the flaw spread and is fading on Previdian's sensors, not that any customer has been breached, and a fading count on five sensors says nothing certain about the rest of the internet.
2. What the attempts were doing. The Hacker News quotes a statement from watchTowr's head of threat intelligence, given to it, that watchTowr is also seeing active exploitation in the wild, that "most attempts" so far have focused on fingerprinting whether an affected product exists and sweeping for common configuration files and sensitive information, and that no post-compromise activity, and no use of credentials taken through the flaw, has been observed. I read that only as quoted by The Hacker News, whose article is not dated on its page, and I did not find watchTowr's own statement. The same article quotes Previdian's chief executive as saying a public scanning template makes mass automated scanning easier and that activity should rise; Previdian's own daily count has since fallen, and I did not find the template itself.
3. Nothing else has moved. CISA's catalogue is still version 2026.10.08 and does not list CVE-2026-21589, 6 days after attempts began (derived). Atlassian's bulletin, re-read this evening, still says its investigation has not found evidence of exploitation in Cloud, and still says that for Data Center it cannot confirm whether any instance was affected, and tells owners to check access logs for evidence of compromise. The fixed builds and the mitigations are as listed above.
What to do differently: nothing in the order of actions above changes, and the falling count is not a reason to relax. File sweeps and configuration reads are how a leaked credential file is found. For any internet-reachable instance since 5 October, search the logs you still hold using Atlassian's own pattern, patch, and where you cannot patch yet cut the instance off from the internet or apply Atlassian's rules. If Jira and Crowd are linked, treat the Crowd application credential as exposed if the logs show the file being read, and rotate it.
What changed after the 9 October update. Sources: Previdian's page (updated 12 October), The Hacker News, Atlassian's bulletin and the CISA catalogue (version 2026.10.08), read on 12 October 2026
| Item | Read on 9 October | Read on 12 October |
|---|---|---|
| Previdian's count | 248 attempts, 42 addresses, 14 countries, four sensors; 27 so far on 9 October | 380 attempts, 58 addresses, 19 countries, five sensors; 6, 107, 108, 86, 54, 15 and 4 a day from 6 to 12 October (12 October partial) |
| What the attempts did | Not stated | watchTowr (via The Hacker News): mostly fingerprinting and sweeping for configuration files and sensitive information; no post-compromise activity observed |
| The CISA catalogue | Version 2026.10.08; not listed | Version 2026.10.08; not listed, 6 days after attempts began (derived) |
| Atlassian's bulletin | No evidence of exploitation in Cloud; cannot confirm for Data Center | Wording unchanged |
| Who has been compromised | No source reports a real instance compromised | Still none reported in any source read; watchTowr says no post-compromise activity seen |
Update, 09:08 BST on 9 October: Previdian's count is 248 attempts from 42 addresses in 14 countries, and CISA's catalogue has been updated without it
At about 09:08 BST on 9 October I re-read Previdian's page for the flaw and fetched the CISA catalogue again. This update prints no address, path, file name or payload.
Previdian's page, updated on 9 October, now shows 248 exploitation attempts from 42 unique addresses in 14 countries, seen by four sensors, against 193, 33 and 11 on the afternoon of 8 October. Its daily chart reads 6 attempts on 6 October, 107 on 7 October, 108 on 8 October and 27 so far on 9 October, a day that was not over when I read it. The first observation is still 6 October. All of it is honeypot traffic: it shows that scanning for the flaw is widespread and still running, not that any customer has been breached.
The CISA catalogue has moved. It is now version 2026.10.08, released on 8 October at 20:09 UTC, with 1,739 entries, five more than the version I last quoted. The five new entries are older flaws in other products, and CVE-2026-21589 is not among them. So the caveat I gave on 8 October, that the catalogue had not changed since 4 October, no longer applies: the catalogue has been updated since attempts began on 6 October, and it still does not list this flaw. Previdian's page says the same. Why CISA has not listed it is not stated; its own criteria are on its site, which I did not re-read today, and my reading is only that honeypot attempts may not meet its bar. The absence says nothing about whether any real instance has been reached, and no source I have read reports one.
What changed after the 15:43 BST update on 8 October. Sources: Previdian's page for the flaw and the CISA catalogue, both read at about 09:08 BST on 9 October
| Item | Read on 9 October | Still not stated |
|---|---|---|
| Honeypot volume | Previdian: 248 attempts, 42 addresses, 14 countries, four sensors; 6, 107, 108 and 27 attempts on 6, 7, 8 and 9 October (9 October incomplete) | Whether any attempt reached a real instance, which of the eight products each attempt targeted, who operates the addresses |
| The CISA catalogue | Version 2026.10.08, 20:09 UTC on 8 October, 1,739 entries; CVE-2026-21589 not listed. Previdian's page also says not in the CISA catalogue | Whether or when CISA will add it, and what evidence it would need |
| Atlassian's bulletin | Not re-read since the 15:43 BST update on 8 October; the wording then was unchanged | Whether Atlassian has seen attempts against Data Center instances |
| Who has been compromised | No source I read reports a real instance compromised, a credential taken or an account created | Any victim, count or actor |
My reading is unchanged and firmer. The count has risen about sixteen-fold since the 17:40 update on 7 October (248 against 15, derived), and the catalogue is not the thing to wait for. For any instance reachable from the internet since 5 October, copy the logs before any node restarts, search them with Atlassian's own pattern, patch, and cut the instance off from the internet or apply Atlassian's rewrite rules where you cannot patch yet. The fixed builds and mitigations in Atlassian's bulletin are as they were.
Update, 15:43 BST on 8 October: Previdian's own page shows 193 attempts from 33 addresses in 11 countries, and CISA's catalogue still does not list it
At about 15:43 BST on 8 October I read Previdian's own page for the flaw, which I had not been able to find before and which is the source of the honeypot counts quoted below, read SecurityWeek's report that cites it, re-read Atlassian's bulletin and fetched the CISA catalogue again. Until now my figures for Previdian came from press accounts of its warning. This update prints no address, path, file name or payload.
Previdian's page, which says it was updated on 8 October, shows its sensors recording 193 exploitation attempts from 33 unique addresses in 11 countries, seen by three sensors. Its daily chart reads 6 attempts on 6 October, 107 on 7 October and 80 on 8 October, a day that was not over when I read it. It dates the first observation to 6 October and the latest to 8 October. SecurityWeek, a little earlier, gave 190 attempts from 32 addresses in 10 countries, so the count was still rising as I read. The 15 attempts from three addresses in the 17:40 update on 7 October were an early reading of this telemetry (my reading: the page does not say so). The page lists the countries but gives addresses, request targets and User-Agents only to paying customers, and I did not look for them.
What changed after the 11:22 BST update on 8 October. Sources: Previdian's page for the flaw, SecurityWeek, Atlassian's bulletin and the CISA catalogue, all read at about 15:43 BST on 8 October
| Item | Reported or read | Still not stated |
|---|---|---|
| Honeypot volume | Previdian's page: 193 attempts, 33 addresses, 11 countries, three sensors; 6, 107 and 80 attempts on 6, 7 and 8 October (8 October incomplete). SecurityWeek: 190, 32 and 10 earlier the same day | Whether any attempt reached a real instance, which of the eight products each attempt targeted, and who operates the addresses |
| Public tooling | Previdian's page records public scanner or proof-of-concept coverage on GitHub on 6 October and a public scanner template detected on 7 October. SecurityWeek ties the start of the attempts to watchTowr's analysis of 6 October | What either does. I have not opened them and this update does not link them |
| The CISA catalogue | Version 2026.10.04, released 4 October at 18:52 UTC, 1,734 entries, unchanged when I fetched it. CVE-2026-21589 is not listed. Previdian's page also says it is not in the CISA catalogue | Whether or when CISA will add it |
| Atlassian's bulletin | Wording unchanged when re-read: Cloud patched, investigation has not found evidence of exploitation; for Data Center, Atlassian says it cannot confirm whether instances were affected and tells customers to check logs for evidence of compromise | Whether Atlassian has seen attempts against Data Center instances |
| Who has been compromised | No source I read, Previdian's page included, reports a real instance compromised, a credential taken or an account created | Any victim, count or actor |
Two points follow. First, the volume is far higher than the 17:40 update could say, roughly a thirteen-fold rise in the count in about a day (193 against 15, derived), and the addresses have risen from three to 33. All of it is sensor traffic: it shows that scanning for the flaw is widespread and automated, not that any customer has been breached. Second, nothing here changes what the sources leave open, and everything here strengthens the advice already given. For any instance reachable from the internet since 5 October, copy the logs before any node restarts, search them with Atlassian's own pattern, patch, and cut the instance off from the internet or apply Atlassian's rewrite rules where you cannot patch yet. The bulletin's fixed builds and mitigations are unchanged.
Update, 11:22 BST on 8 October: a commercial catalogue lists it as exploited, and CISA's does not
At about 11:22 BST on 8 October I read Infosecurity Magazine's report, re-read Atlassian's bulletin and fetched the CISA catalogue again. Infosecurity says VulnCheck, a vulnerability-intelligence company, added CVE-2026-21589 to its own known-exploited list on 7 October, observing exploitation activity that targets Bamboo Data Center, and that VulnCheck's dashboard links Previdian, the firm behind the honeypot reports in the update below, as a source. VulnCheck's page asked me to sign in, so I could not read it, and this is Infosecurity's account of it. Infosecurity does not say whether the activity VulnCheck observed was against a real instance or a honeypot, and nothing I read says any real instance was compromised. This update prints no path, address, file name or payload.
What changed after the 17:40 BST update on 7 October. Sources: Infosecurity Magazine; Atlassian's bulletin and the CISA catalogue, both read at about 11:22 BST on 8 October
| Item | Reported or read | Still not stated |
|---|---|---|
| A commercial catalogue | Per Infosecurity, VulnCheck listed the flaw as exploited on 7 October, with activity aimed at Bamboo Data Center, citing Previdian | What VulnCheck saw, from which sensors, whether any real instance was reached. I could not read VulnCheck's page |
| The CISA catalogue | Version 2026.10.04, released 4 October at 18:52 UTC, 1,734 entries, unchanged: more than three and a half days old when read (derived). CVE-2026-21589 is not listed. Infosecurity says the same | Whether or when CISA will add it. A catalogue that has not changed since 4 October cannot say anything about this flaw either way |
| Atlassian's bulletin | Wording unchanged when re-read: Cloud, no evidence of exploitation; Data Center, Atlassian cannot confirm whether instances were affected | Whether Atlassian has seen attempts against Data Center instances |
| Who has been compromised | No source I read reports a real instance compromised, a credential taken or an account created | Any victim, count or actor |
Two points follow. First, a listing by VulnCheck and a listing by CISA are different things. The CISA catalogue is the list the NCSC's active exploitation guidance refers to; VulnCheck's is a commercial list with its own criteria, which the sources do not set out. A listing there is a second party saying exploitation has been seen, and, since Infosecurity says VulnCheck cites Previdian, it appears to rest on the same telemetry the update below describes, so it is more outlets for one source and not an independent confirmation. Second, my reading in the update below stands and is firmer: for any instance reachable from the internet since 5 October, copy the logs before any node restarts, search them with Atlassian's own pattern, then patch. The fixed builds and the mitigations in Atlassian's bulletin are unchanged.
Update, 17:40 BST on 7 October: attempts against the flaw are reported, and it is still not in the KEV catalogue
The article below was written on the morning of 6 October, when no attempt to use the flaw had been reported. At about 17:40 BST on 7 October I read The Hacker News, BleepingComputer, Help Net Security and the SANS Internet Storm Center diary, read watchTowr Labs' own post (listed in the sources) for its date and claims, re-read Atlassian's bulletin and fetched the CISA catalogue again. This update prints no request, address, file name or payload. Everything about the technique below is the researchers' account as I read it. I have not tested any of it.
What was published. watchTowr Labs, whose site dates its post 6 October, the day after Atlassian's bulletin, traced the cause to a web-resource library shared by the affected products, and showed requests that read protected application files in Jira, Confluence and Bitbucket, according to its post and to Help Net. For Crowd, it says a configuration file under the web application holds the application's name and password in plain text. With those, an attacker who can reach Crowd can create users and raise one to Jira administrator. watchTowr adds that Crowd can be set to accept only listed IP addresses, which would make that route much harder.
What was seen. Previdian, a threat-intelligence firm, says its honeypots recorded 15 attempts from three addresses in Japan and the United States, the first within two hours of watchTowr's post (The Hacker News; BleepingComputer quotes Previdian on the two hours). Help Net says Previdian's warning came late on Tuesday 6 October. BleepingComputer, writing on 7 October, says "earlier today". I did not find Previdian's own post, so the count and the day come from those three outlets. Previdian also says a template for automated scanners has been released and expects activity to rise. Separately, the SANS Internet Storm Center's dean of research writes that its honeypots saw attempts from 6 October, from 13 addresses on one cloud provider, aimed at routes for Jira, Bitbucket and Confluence, and that he believes one actor is behind them. The two sets of addresses do not overlap (my comparison of the lists the two sources print), so they come from different networks of sensors and say nothing about whether they share an operator.
What the evidence is. Headlines say the flaw is "exploited". What the four reports describe is attempts logged by honeypots, machines that exist to be probed. None reports a real instance whose files were read, a credential taken or an account created, and none says how many internet-facing instances exist. That is a statement about these four reports, not about the internet.
What changed after the article was written on the morning of 6 October. Sources: The Hacker News, BleepingComputer, Help Net Security, SANS Internet Storm Center, watchTowr Labs, Atlassian's bulletin and the CISA catalogue, all read at about 17:40 BST on 7 October
| Item | Reported or read | Still not stated |
|---|---|---|
| Exploitation | Honeypot attempts: 15 from three addresses (Previdian) and 13 addresses from 6 October (SANS Internet Storm Center). The Hacker News and BleepingComputer both say exploitation has begun | A successful read of a real instance's files, any victim, any actor, any count of exposed instances |
| Public detail | watchTowr's post of 6 October shows the technique and a route from file read to Jira administrator through Crowd. Previdian says a scanner template is out. The sources differ on whether a ready-made exploit is public: Help Net says watchTowr built one and did not publish it, Previdian says "public PoC" | Whether a ready-to-run exploit is public beyond what the write-up shows. I did not look for one |
| Atlassian's bulletin | Re-read at about 17:40 BST. It still says its investigation has not found evidence of exploitation in Cloud, and that it cannot confirm whether any Data Center instance was affected. Fixed builds and mitigations are unchanged | Whether Atlassian has seen attempts against Data Center instances. The bulletin does not say |
| CISA KEV catalogue | Version 2026.10.04, released 4 October 18:52 UTC, 1,734 entries, unchanged. CVE-2026-21589 is not listed. No entry of any kind has been added since the flaw was published | Whether or when CISA will add it |
What this changes below. First, the article said the NCSC's rule on checking for compromise before updating was not triggered, because exploitation was not established. Attempts are now documented, and the honeypot reports cannot say whether anyone tried a real instance earlier. My reading is that any instance reachable from the internet since 5 October should be treated as if the rule applies: copy the logs before any node restarts, search them, then patch. That is judgement, mine, and not NCSC wording about this flaw.
Second, the log-search advice below is unchanged and fits what is reported. The attempts the SANS diary describes use the double-colon spelling, which is one of the spellings Atlassian's own search advice names, so the bulletin's check is the test. Third, the article's judgement that the exact-name condition raises the bar only for files an administrator added looks borne out: the diary says the file used as a test exists in every Tomcat web application.
Fourth, the article closed by asking which file under the web directory you would least want a stranger to read. For Crowd-integrated deployments the reports now name one: a configuration file that holds the application's credentials. Ask who can reach Crowd and whether it accepts only listed addresses. Rotating that credential after patching is my judgement, not Atlassian's advice, which is to patch, mitigate and check for evidence of compromise.
Nineteen fixed builds, and a short runway under seven of them
Atlassian's bulletin for CVE-2026-21589 names eight products and lists 19 fixed builds between them. Counted against Atlassian's own end-of-support policy page, seven of those 19 builds sit on release lines whose support ends between 3 and 30 December 2026, which is 58 to 85 days from Tuesday 6 October. Three more sit on lines that page does not list, and nine run into 2027 or 2028. These counts are derived by this site from the two Atlassian pages, and the map further down draws every one of them.
That is a fact about how long this week's patch will carry you. It says nothing about exploitation, and attempts have been reported since this article first appeared (see the update above). When first published, it said: Atlassian says its investigation found no evidence of exploitation in Cloud, says it cannot confirm whether any Data Center instance was affected, and the CISA catalogue of known exploited flaws did not list the CVE when read. On 7 October at 17:40 BST the bulletin's wording was the same and the catalogue still did not list the CVE. Nor does the count say which files a stranger could read on your instance, how many instances are exposed to the internet, or whether the flaw has anything to do with earlier Atlassian bugs. The primaries are silent on all three, and so is this article.
State of each source when read on the morning of Tuesday 6 October 2026, UTC. Any of these can change; the times say when each was last read.
| Source | State when read | Read at (UTC) |
|---|---|---|
| Atlassian bulletin | Advisory release date 5 Oct 2026; page last modified 5 Oct 2026; no revision table on the page | 05:59, same as 05:36 |
| Atlassian public tickets (eight) | All eight Published and Fixed, resolved 5 Oct between 21:04 and 21:05 UTC | 05:59, same as 05:38 |
| CVE.org record | PUBLISHED 5 Oct 21:30 UTC; last updated the same minute | 05:59, same as 05:35 |
| NVD record | Received, 5 Oct 22:16 UTC; no NVD or CISA-ADP score, no CWE | 05:59, same as 05:35 |
| CISA KEV catalogue | Version 2026.10.04, released 4 Oct 18:52 UTC, 1,734 entries; CVE-2026-21589 not listed | 05:59, same as 05:37 |
What the primaries say, and what they leave out
Everything in the Stated column below is on Atlassian's bulletin, its eight public tickets, the CVE.org record or the NVD record. The Not stated column is what none of them states. Reading the second as if it were the first is how a short advisory turns into a long rumour.
Stated and not stated, from the bulletin, the eight Atlassian tickets, the CVE.org record, the NVD record and the CISA KEV catalogue, read 6 October 2026.
| Topic | Stated | Not stated |
|---|---|---|
| Reach | Unauthenticated, over the network, no user interaction. Atlassian says to restrict internet access even where instances use authentication | How many instances are exposed. No scan figure is quoted here |
| What is read | Specific files in the web application root. The exact name and path must be known. No directory listing | Which files, or which configurations put sensitive ones there |
| Exploitation | Cloud: patched, no evidence of exploitation found. Data Center: Atlassian cannot confirm whether instances were affected. Bulletin wording unchanged when re-read on 7 October | In these sources as read on 6 October: any exploitation of Data Center, any public exploit, who found the flaw, or when. Honeypot reports by others since then are in the update above |
| Scope | All versions of eight Data Center products | Which shared component is at fault, or why eight products share it |
| Fix | 19 builds; release notes date 16 of them 5 Oct 2026 and Bitbucket's three 6 Oct 2026 | A fixed build in every supported line: 34 supported lines have none listed |
| Score | 9.3 Critical, CVSS 4.0, Atlassian's own assessment | An NVD, CISA-ADP or CVSS 3 score, or a CWE identifier |
| Server editions | The CVE record lists eight Server products as affected | Anything about Server in the bulletin; most have no fix listed |
Two dates are worth separating from discovery. The earliest of the eight public tickets was created on Friday 2 October at 04:26 UTC, three days and 16 hours before Atlassian resolved it, which shows the ticket existed, not when the flaw was found. The CVE identifier was reserved on 1 January 2026, 277 days before publication, which shows Atlassian reserves identifiers in advance, not that it knew of the flaw in January. By the dates on the bulletin the gap between Atlassian's email on Monday 5 October, as The Register reports it, and the bulletin is zero days.
The fix map: which line, which build, until when
Atlassian's advice is blunt: it "recommends patching to the fixed LTS version or later". Most fixed builds are on long-term-support lines. The exceptions are Bitbucket 10.5, which the policy page does not list, and three Crowd lines not marked LTS. Atlassian's security bug fix policy says critical fixes go to the latest feature release and to supported LTS releases. Support for each line runs two years from its first release, according to the policy page, last modified 22 September 2026. Put the bulletin and the policy page side by side and the runway is uneven.
Four points follow from the map, each checkable on the two Atlassian pages.
Most supported lines have no fixed build. The policy page lists 34 other release lines that are still in support after 6 October and have no fixed build in the bulletin (derived: Bamboo 3, Bitbucket 7, Confluence 5, Crowd 1, Jira Software 9, Jira Service Management 9). A Confluence 9.5 site, supported to 4 June 2027, has no fixed 9.5 build to install. It has to move to a listed build or later. Confluence 9.1 left support on 3 October, three days before this article, and is on no fix path at all.
The Jira 9.12 and Service Management 5.12 lines are an open question. The Jira Software 9.12 release notes give 29 November 2025 as that line's end-of-life date, the policy page does not list 9.12 or 5.12, and Atlassian released 9.12.40 on 5 October 2026, 310 days after the stated date. The sources do not say whether support was extended. Anyone on those lines should ask Atlassian rather than assume.
Older versions are not cleared. Each of the eight tickets says versions that have reached end of life "may also be affected". That is a statement of untested scope, not of safety.
Server editions sit outside the bulletin. The CVE record lists eight Server products as affected, and Atlassian's Server end-of-support notice put the end of support at 15 February 2024, 963 days before this bulletin. The bulletin itself covers Data Center only.
A 9.3 built on what lies downstream
Atlassian scores the flaw 9.3, Critical, with the vector CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:N/VA:N/SC:H/SI:H/SA:H, "per our internal assessment". It adds: "This is our assessment, and you should evaluate its applicability to your own IT environment." The score is Atlassian's alone. NVD had assigned none when read, and CISA-ADP had added none.
The vector says two different things. For the vulnerable system, confidentiality impact is high and integrity and availability impacts are none, which is what a read-only flaw should score. For the subsequent systems, the three metrics SC, SI and SA are all high. This site recomputed the score from the vector with two independent open-source implementations of the CVSS 4.0 specification: both return 9.3. Setting only SC, SI and SA to none returns 8.7, High, in both. So the whole step from High to Critical rests on the assumption that a file read leads to harm elsewhere. Atlassian's own caveat points the same way: "In some configurations, there may be sensitive files present that increase your risk."
That is not a criticism. A vendor cannot see your web root, so it plausibly scores a worst reasonable case; the bulletin does not explain its choice of values. The point is what the number cannot do: it cannot tell you which case you are in. Both 9.3 and 8.7 are above the 7.0 line that Cyber Essentials applies to CVSS 3 scores, though the versions differ, so the arithmetic changes the label and not the clock. The site has looked before at how two scoring versions of one flaw land in different bands.
The friendly-name fallacy runs both ways. "File access" sounds milder than remote code execution, and The Register is right that, in Atlassian's words, exploitation "requires prior knowledge of the target file's exact name and path" and directories cannot be listed. Those are stated preconditions and are true to the primary. What the primaries do not say is how hard the name and path are to know. This is judgement, labelled as such: Atlassian publishes its software, so the layout of a default install can be learned by anyone, and the precondition raises the bar for files an administrator added, not for files every install has.
Two further cautions on labels. The CVE record's problem type is free text, "Path Traversal (Arbitrary Read/Write)", with no CWE identifier, while the vector scores no integrity impact on the vulnerable system. Treat the flaw as read until Atlassian says otherwise. And Atlassian's block rule and detection advice both concern two dots next to a path separator, which suggests, as inference only, that the fault is in how the request path is handled. The bulletin names no component.
For context, not as a link: CISA's catalogue holds 13 Atlassian entries, three of them described as path traversal or arbitrary file read (CVE-2019-3398, CVE-2021-26085 and CVE-2021-26086). The class has been exploited in Atlassian products before. Nothing in the primaries connects any of those flaws to this one.
Where Atlassian's own records disagree
The bulletin is Atlassian's page. The CVE record is Atlassian's submission to the CVE programme. Both appeared on 5 October, the tickets and the record about 25 minutes apart, and they do not agree. For anyone who patches from a vulnerability scanner, a ticketing feed or a CVE-record parser, the record is the version they will see.
Disagreements among Atlassian's bulletin, its public tickets, its release notes, its download feed and its CVE.org record, read 6 October 2026.
| Item | One Atlassian source says | Another says |
|---|---|---|
| Crowd 7.1 fixed build | Bulletin and the ticket's fix-version field: 7.1.7 | CVE record: 7.1.1, released 27 Nov 2025, 312 days before the bulletin. Ticket text: 7.1.6, which the same ticket lists as affected |
| Bamboo 10.2 fixed build | Bulletin and CVE description: 10.2.24 | CVE affected-version data: 10.2.4 and later. One third-party tracker page repeats it |
| Release dates of the fixed builds | Release notes: 5 Oct 2026 for 16 builds, 6 Oct for Bitbucket's three | Download feed: 2 Oct for two Jira builds, 6 Oct for four Crowd builds. Ticket version data differs again |
| How far back | Bulletin: all versions | CVE record: introduced in builds from Jira Software 7.1.0 and Service Management 3.1.0 (10 Feb 2016) to Bamboo 7.0.1 (10 Mar 2020); nothing for Crucible or Fisheye |
| Editions | Bulletin: Data Center only | CVE record: eight Server products also affected |
| What the flaw allows | Bulletin and vector: file access; no integrity impact on the vulnerable system | CVE problem type: Read/Write |
The practical rule is short. Use the bulletin's build numbers, confirm the build you install is at or above the number for your line, and check the release notes page for that line. Do not let a scanner mark Crowd 7.1.1 or Bamboo 10.2.4 as fixed. The oldest line the record names was released more than ten and a half years before disclosure (3,890 days from 10 February 2016 to 5 October 2026, derived), but the record's start versions are the CVE submission's claim, not the bulletin's.
What discontinuing Data Center does and does not cover
The Register says Atlassian "decided to discontinue its datacenter software", and a reader could take from that one end date for all eight affected products. Atlassian's own end-of-life page says something narrower, and this article follows the vendor page.
Data Center end-of-life dates for the eight affected products, from Atlassian's end-of-life page and support pages read 6 October 2026. Day counts derived from 6 October 2026.
| Product | Atlassian's stated date | Days |
|---|---|---|
| Jira Software, Jira Service Management, Confluence and Crowd Data Center | End of life 28 Mar 2029 at 23:59 PST, then read-only | 904 |
| Bitbucket Data Center and Bamboo Data Center | "Will not end of life", via a Bitbucket Hybrid License | no date |
| Crucible and Fisheye | Support ends 15 May 2028; no new sales since 13 May 2025 | 587 |
Through 28 March 2029 Atlassian promises technical support and security fixes for critical vulnerabilities for the affected Data Center products, and new customers have been unable to buy Data Center since 30 March 2026, 190 days ago. Existing customers can expand until 30 March 2028. After the end date the instance is read-only, and Atlassian says: "We strongly advise against keeping your product running in a read-only mode while connected to the internet."
Two UK points. First, Cyber Essentials requires in-scope software to be licensed and supported, and defines support as a vendor commitment with a stated end date; Atlassian's dates are those dates. Second, for bodies on an April to March financial year, the end-of-life date, 28 March 2029, is three days before that year closes on 31 March 2029, and the last day to expand, 30 March 2028, is one day before the year closing on 31 March 2028. That is judgement, not a source claim: a migration first funded in the 2028 to 2029 year starts in the last year of support.
The UK clocks, and what the NCSC does and does not say
Two published clocks apply to a self-hosted instance, and both start at release. The bulletin is dated Monday 5 October, and the release notes date 16 of the 19 fixed builds to the same day; Bitbucket's three say 6 October, a day later, with no time zone given.
Patching clocks from a 5 October 2026 release. NCSC vulnerability management guidance (published 12 Feb 2024, reviewed 1 May 2026, version 2.1) and Cyber Essentials Requirements for IT Infrastructure v3.3 (April 2026), read 6 October 2026.
| Clock | Source | Counted from 5 Oct |
|---|---|---|
| 5 days, internet-facing services and software | NCSC best practice | Saturday 10 Oct 2026 |
| 14 days, vendor-rated critical or high, or CVSS 3 base 7 or above | Cyber Essentials v3.3 | Monday 19 Oct 2026 |
| 14 days, internal and air-gapped services | NCSC best practice | Monday 19 Oct 2026 |
Cyber Essentials counts "critical or high risk" by a CVSS 3 base score of 7 or more, or by the vendor's own wording. Atlassian publishes only a CVSS 4.0 vector and calls the flaw Critical, so the vendor-wording route applies; that is this site's reading. If the 2 October date in Atlassian's download feed for two Jira builds is the one that counts, the 14 days end on 16 October. The NCSC guidance also says that for an actively exploited internet-facing flaw you should check for compromise before updating. When this article first appeared exploitation was not established, so that rule was not triggered. Honeypot attempts have since been reported, and my reading, given in the update above, is that for an instance reachable from the internet the cautious course is to copy the logs before any node restarts and check them for compromise.
What the NCSC has said about this flaw: nothing found. Keyword searches of ncsc.gov.uk at about 06:57 BST on 6 October returned six items for "Atlassian" and three for "Confluence", all general guidance or research, and nothing about this CVE. Nothing found is specific to collaboration or developer tooling either. The closest guidance is general. The NCSC's 27 August 2026 alert on internet-exposed systems is operational-technology focused, but tells other organisations to keep an accurate inventory of internet-facing systems, apply vendor updates promptly, retire end-of-life equipment and register for the free Early Warning service. Early Warning works from the IP addresses and domains you register; whether it will flag this flaw is not stated. The NCSC's January 2025 paper on 'unforgivable' vulnerabilities lists CWE-22 path traversal and scores input validation, one of its mitigations, as easy, but says the scoring assumes mitigations are built in from the start and not retrofitted. With no code-level cause published, this article makes no forgivable or unforgivable call.
Who is exposed in the UK is not in the sources. There is no UK install count. Atlassian's own end-of-life page says "source code and CI/CD pipelines are mission-critical" and some organisations "may need to keep them self-managed". Judgement: that profile fits public bodies, universities, NHS trusts and defence suppliers that keep source code or tickets in-house, but the sources name no sector. The first item in Atlassian's shared-responsibility list is "Operating Atlassian software on private networks".
What sits in a web root is the open question. The bulletin does not list files. Atlassian's Confluence documentation puts the start-up configuration file, which holds database connection settings, in the home directory and not the installation directory. In that layout those settings are not under the web application. "In some configurations" is then the risk: a layout where someone has put configuration, backups, exports or scripts under the application directory. Whether yours is one of them is a check, not a guess.
What to do, in the order worth doing it
Take this with you
Order of work
- Find every Atlassian instance and version: Bitbucket, Confluence, Jira Software, Jira Service Management, Bamboo, Crowd, Crucible and Fisheye, including test, staging, inherited and old Server-era installs. Use licence records, DNS and reverse proxy configuration, and ask who runs Crucible or Fisheye, which are easy to forget.
- Take each internet-facing instance off the internet, or put it behind a VPN or single sign-on at the proxy, until it is patched. Atlassian says this applies even where instances use authentication.
- For anything you cannot patch today, apply the block rule in the bulletin for your product, at the proxy or firewall, in Tomcat or in Bitbucket's rewrite file. Back up first, test it, and keep the patch date.
- Patch to a build at or above the bulletin's number for your line, or move to a newer listed line. If your line has no fixed build, plan the move now. Use the bulletin's numbers, not the CVE record's, for Crowd and Bamboo.
- Before restarting any node, copy web server, proxy and application access logs off the host.
- Search the logs for as far back as you keep them. Atlassian's own advice is to decode each request line, up to two passes, and look for two dots next to a path separator, or to search raw lines with the bulletin's pattern. Look for successful responses as well as attempts, and for one source sending many differently shaped requests. The honeypot reports date attempts to 6 and 7 October, but they say nothing about real instances before then, so the window is every retained log.
- Rotate any secret that sat in a file under a web application directory, or that you cannot show did not. This is judgement: Atlassian does not tell you to rotate anything, and rotating costs less than not knowing.
- Put the next upgrade in the plan: any instance patched to a line that ends support in December 2026 already has a date. Cyber Essentials needs supported software.
- Register the organisation for NCSC Early Warning, and subscribe a named person to Atlassian's advisory emails so the next notice has an owner.
A block rule is a stopgap, not a fix. The site's earlier briefing One letter past the WAF covers what happens when a rule matches one spelling of a path and the attacker uses another. Atlassian's pattern is written to cover encoded forms, but only a patched build removes the dependence on any rule.
The question the score cannot answer
Atlassian has done the part a vendor can do: it published fixed builds, a block rule, detection advice and a plain statement that it cannot confirm whether any instance was affected. The score, the clocks and the end-of-life dates all assume something the sources cannot see, which is what sits under your web root and who last checked.
So the question is not whether this flaw is a 9.3 or an 8.7. For each Atlassian instance you run, can you name the one file under its web application directory that you would least want a stranger to read, and show who checked that it is not there?
Key facts
Sources
- PrimarySecurity bulletin for CVE-2026-21589, advisory release date 5 Oct 2026: affected products, fixed versions, CVSS 4.0 vector, mitigations, detection advice, Cloud statementAtlassianaccessed 2026-10-08
- PrimaryCVE.org record: published 5 Oct 2026 21:30 UTC, affected and unaffected version data, introduced-in versions, problem type, 4.0 vectorCVE Program (CNA: Atlassian)accessed 2026-10-06
- PrimaryNVD record: status Received, published 5 Oct 2026 22:16 UTC, Atlassian's score only, no CWE, no CISA-ADP dataNIST National Vulnerability Databaseaccessed 2026-10-06
- PrimaryKnown Exploited Vulnerabilities catalogue, version 2026.10.04, 1,734 entries; CVE-2026-21589 absent; 13 Atlassian entriesCISAaccessed 2026-10-08
- PrimaryPublic ticket for Crowd: fix-version field 7.1.7, description table 7.1.6, creation and resolution times; the seven sibling tickets were read the same wayAtlassianaccessed 2026-10-06
- PrimaryPublic ticket for Confluence: earliest created of the eight, 2 Oct 2026 04:26 UTC; end-of-life versions may also be affectedAtlassianaccessed 2026-10-06
- PrimaryData Center end of life: 28 Mar 2029 for Jira Software, Jira Service Management, Confluence and Crowd; Bitbucket and Bamboo not ending; sales and support datesAtlassianaccessed 2026-10-06
- PrimaryAtlassian End of Support Policy, last modified 22 Sep 2026: end-of-support date of every release line; Fisheye and Crucible support to 15 May 2028Atlassianaccessed 2026-10-06
- PrimarySecurity Bug Fix Policy: critical fixes to the latest feature release and supported LTS releases; shared responsibility listAtlassianaccessed 2026-10-06
- PrimaryServer end of support: 15 Feb 2024, and Fisheye and Crucible support end 15 May 2028Atlassianaccessed 2026-10-06
- PrimaryJira Software 9.12 release notes: end-of-life date 29 Nov 2025; 9.12.40 released 5 Oct 2026Atlassianaccessed 2026-10-06
- PrimaryCrowd 7.1 release notes: 7.1.1 released 27 Nov 2025; 7.1.7 released 5 Oct 2026Atlassianaccessed 2026-10-06
- PrimaryBitbucket Data Center 9.4 release notes: 9.4.26 released 6 Oct 2026 (10.2 and 10.5 notes read the same way)Atlassianaccessed 2026-10-06
- PrimaryDownload feed: Jira Software 9.12.40 and 10.3.26 dated 2 Oct 2026, 11.3.12 dated 5 Oct 2026Atlassianaccessed 2026-10-06
- PrimaryConfluence documentation: installation directory versus home directory and what the start-up configuration file holds, last modified 10 Dec 2024Atlassianaccessed 2026-10-06
- PrimaryVulnerability management guidance, update by default: 5 days for internet-facing services, 14 days internal; published 12 Feb 2024, reviewed 1 May 2026NCSCaccessed 2026-10-06
- PrimaryCyber Essentials Requirements for IT Infrastructure v3.3, April 2026: 14-day security update rule, licensed and supported software definitionNCSCaccessed 2026-10-06
- PrimaryAlert of 27 Aug 2026 on internet-exposed systems and edge devices: inventory, vendor updates, retire end-of-life, Early WarningNCSCaccessed 2026-10-06
- PrimaryResearch paper of 28 Jan 2025 on forgivable and unforgivable vulnerabilities: CWE-22 mitigations and the retrofit caveatNCSCaccessed 2026-10-06
- PrimaryEarly Warning service: free alerts for UK organisations registered by IP address and domainNCSCaccessed 2026-10-06
- PrimarywatchTowr Labs post dated 6 Oct 2026: cause in a shared web-resource library, file read in Jira, Confluence and Bitbucket, escalation through Crowd, detection artefact generatorwatchTowr Labsaccessed 2026-10-07
- PrimaryCVE-2026-21589 Exploitation Observed, page updated 8 Oct 2026: 193 attempts, 33 addresses, 11 countries, three sensors, daily counts 6 / 107 / 80, public scanner template detected 7 Oct, not in CISA KEV; read again 9 Oct: 248 attempts, 42 addresses, 14 countries; read again 12 Oct: 380 attempts, 58 addresses, 19 countries, five sensorsPrevidianaccessed 2026-10-12
- Reported byNews report of 6 Oct 2026, 05:20 UTC, used as the pointer: Monday email, 'Action required', filename precondition, cloud and Data Center framingThe Registeraccessed 2026-10-06
- Reported byThird-party tracker page repeating the CVE record's Bamboo '10.2.4 and later' figure; not relied on for any other factSecurityOnlineaccessed 2026-10-06
- Reported byDiary 33406, 7 Oct 2026: honeypot attempts from 6 Oct, 13 addresses on one cloud provider, routes for Jira, Bitbucket and ConfluenceSANS Internet Storm Centeraccessed 2026-10-07
- Reported byExploitation attempts against critical Atlassian flaw have begun, 7 Oct 2026: Previdian warning, watchTowr summary, exploit not publishedHelp Net Securityaccessed 2026-10-07
- Reported byAtlassian Data Center Flaw Draws Exploitation Attempts Within Two Hours of Public Details, 7 Oct 2026: 15 attempts from three addressesThe Hacker Newsaccessed 2026-10-07
- Reported byHackers exploit critical Atlassian flaw after public PoC release, 7 Oct 2026: Previdian honeypot detections, scanner template, Crowd caveatBleepingComputeraccessed 2026-10-07
- Reported byCritical Flaw in Multiple Atlassian Products Exploited in the Wild, 8 Oct 2026: VulnCheck listing of 7 Oct, Bamboo Data Center, PrevidianInfosecurity Magazineaccessed 2026-10-08
- Reported byAttackers Target Critical Atlassian Vulnerability Within Hours of PoC Publication, 8 Oct 2026: Previdian's 190 attempts from 32 addresses in 10 countries, watchTowr's Crowd routeSecurityWeekaccessed 2026-10-08


