OFFSEC · Offensive Security

A 9.3 That Needs You to Know the Filename: Atlassian’s Web-Root File Read

Atlassian patched one unauthenticated file-access flaw across eight self-hosted products on 5 October. It cannot list a directory, which makes the score look strange until you consider what organisations leave in a web root.

What Atlassian published

On 5 October Atlassian released an advisory for CVE-2026-21589, an arbitrary file access vulnerability carrying a CVSS 4.0 score of 9.3 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” [1]. The description is short: the flaw “allows an unauthenticated attacker to access specific files within the web application root directory” [1][3].

The affected list is unusually wide because the defect sits in shared platform code. All versions of Bitbucket, Confluence, Jira Software, Jira Service Management, Bamboo and Crowd Data Center are affected, along with Crucible and Fisheye [1][6]. Atlassian states that “Affected Atlassian Cloud products have been patched, and our investigation has not found evidence of exploitation” [1].

Fixed releases are per-product rather than a single version: Bitbucket 9.4.26, 10.2.8 and 10.5.1; Confluence 9.2.26 and 10.2.19; Jira Software and Jira Service Management 9.12.40, 5.12.40, 10.3.26 and 11.3.12; Bamboo 10.2.24 and 12.1.12; Crowd 6.3.7, 7.0.3, 7.1.7 and 7.2.4; Crucible and Fisheye 4.9.15 [1][4][6]. An estate running several of these products needs several upgrade windows, which is why this takes weeks rather than an afternoon.

The constraint that makes this interesting

Atlassian is specific about what the flaw cannot do: “Exploitation requires prior knowledge of the target file’s exact name and path; this vulnerability does not allow attackers to enumerate or list directory contents” [1][4]. watchTowr records the same limit — “attackers need to already know the exact file name and path” — and notes the apparent tension with a 9.3 rating [2].

That constraint is real but weaker than it sounds, for two reasons. The first is that the products are off-the-shelf. Every file that ships with Confluence or Jira sits at a path an attacker can read from a public download, so “prior knowledge” is a documentation lookup for anything the vendor placed there. The constraint bites only on files an administrator added with a name nobody else can guess.

The second is what administrators do add. Atlassian’s own framing, as watchTowr reads it, is that “some installations keep sensitive files inside that same directory” [2] — the familiar inventory of configuration files, database credentials, API keys and application secrets that accumulate in a web root over a decade of upgrades and migrations [2]. MITRE’s weakness class for this is CWE-552, where “the product makes files or directories accessible to unauthorized actors, even though they should not be” [7].

So the honest severity assessment is not vendor-side at all. It depends on a question only the operator can answer: has anything sensitive ever been written into, or left behind in, the web application root of these instances. For an estate with a long history, the answer is often yes and nobody remembers what.

Why 9.3

The score repays a moment’s attention, because it is a good illustration of how CVSS 4.0 behaves. The vector records no privileges required, no user interaction, network attack vector and low attack complexity, with no attack requirements — and then high impact on all three subsequent-system metrics, SC, SI and SA, while the vulnerable system’s own integrity and availability impacts are rated none [1].

Those subsequent-system metrics are the engine of the 9.3. The specification defines them as measuring impacts that occur outside the vulnerable component, reflecting consequences “to the ‘things that suffer the impact'” and the “downstream impact on what is formally called the ‘subsequent system(s)'” [8]. In other words: the file read is modest, but credentials recovered from a web root are used somewhere else, and that is where the harm lands.

It is also worth noting what Attack Requirements is for, since its absence here contributes to the score. The specification describes AT as capturing “the prerequisite deployment and execution conditions or variables” that “emerge naturally as a consequence of the deployment and execution of the vulnerable system” [8]. Needing to know a filename is arguably that kind of condition; the scorers did not treat it as one.

The specification’s own caution is the fairest place to leave this: consumers “may use CVSS information as input to an organizational vulnerability management process that also considers factors that are not part of CVSS”, and matters such as “regulatory requirements, number of customers impacted, monetary losses due to a breach” are “outside the scope of CVSS” [8]. A 9.3 is a statement about severity in the abstract. Whether it is a 9.3 for you is a different and shorter calculation: look in your web root.

The mitigations, and what each actually does

Three temporary measures are published: a web application firewall rule filtering traversal patterns, a rule for the Apache Tomcat rewrite valve, and a URL rewrite for Bitbucket [1][4][9]. The Tomcat option is the one most Atlassian administrators can apply without new infrastructure, since the application server is already there. Tomcat documents the valve as implementing “URL rewrite functionality in a way that is very similar to mod_rewrite from Apache HTTP Server”, configured with a rewrite.config file at Host or Context level, and able to “Forbid or mark URLs as gone” [9].

All three are pattern-matching on request paths, which puts them in the same category as any traversal filter: useful, and not a boundary. watchTowr’s guidance is that patching is strongly preferred, with “restrict public internet access to affected instances” and the vendor’s firewall rule offered as temporary measures that “don’t replace patching” [2]. Atlassian’s own line to administrators who cannot act at once is that instances “accessible to the public internet… should be restricted from external network access until you can take action” [6][10].

What to look for, and what is not known

Detection here is cheap because the request shape is distinctive. watchTowr suggests searching access logs for “requests containing two dots next to a slash, backslash, or double colon” [2]. That will produce noise from scanners, which is itself informative: a sudden rise against these products is a reasonable proxy for the window in which attention arrives.

On exposure, the published record is thin and should be left thin. Nobody has put a credible number on internet-facing instances, and three of the write-ups that would normally carry one do not [4][6][10]. As of 6 October, “no exploitation in the wild has been reported” and the flaw was not listed in the US exploited-vulnerability catalogue [2]; SOC Prime adds that Atlassian “has not confirmed whether individual customer installations were exploited” [5]. An absence of confirmed exploitation of self-hosted instances is not a finding about those instances; it is a statement about what the vendor can see, which for software a customer runs themselves is very little.

What to do

  • Inventory which of the eight products you run self-hosted, and schedule the per-product upgrades; there is no single version that covers the estate [1][4][6].
  • Before or during that work, list what is actually in each instance’s web application root and remove anything holding a credential, key or connection string [2][7].
  • If an upgrade window is weeks away, apply the Tomcat rewrite-valve or firewall rule and record it as temporary, not as remediation [1][2][9].
  • Restrict public internet reachability of these instances in the interim, as the vendor advises [6][10].
  • Grep access logs for traversal sequences against these hosts, and treat a rise in scanning as your timeline rather than the advisory date [2].
  • If anything sensitive was in the web root, rotate it rather than assuming the absence of reported exploitation covers your instance [5][7].

The general lesson is about where severity is decided. A vendor can tell you the mechanism and bound the capability — read a known file, no listing — but it cannot tell you what is in your web root, and that is the entire difference between a nuisance and a credential breach. The useful response to an advisory of this shape is not a debate about the score. It is twenty minutes with a directory listing.

Sources

Every R3KONX article cites its primary material. 10 sources, in order of first citation. Links open the original publication.

  1. CVE-2026-21589: Arbitrary File Access Vulnerability impacts Multiple Products Atlassian · 2026-10-05
  2. Atlassian Data Center and Server Products Vulnerability FAQ: CVE-2026-21589 watchTowr · 2026-10-06
  3. Atlassian Patches Critical Vulnerabilities in Jira, Confluence, Bitbucket, and Five More Products Cyber Security News · 2026-10-06
  4. Atlassian warns of critical file-access flaw in Jira, Confluence BleepingComputer · 2026-10-06
  5. CVE-2026-21589: Critical Atlassian Vulnerability Exposes Sensitive Files Across Eight Products SOC Prime · 2026-10-06
  6. Atlassian urges immediate patching of critical Data Center file access vulnerability (CVE-2026-21589) Help Net Security · 2026-10-06
  7. CWE-552: Files or Directories Accessible to External Parties MITRE · 2026-10-07
  8. Common Vulnerability Scoring System version 4.0: Specification Document FIRST · 2026-10-07
  9. Apache Tomcat 9 Configuration Reference: The Rewrite Valve The Apache Software Foundation · 2026-10-07
  10. Atlassian Warns of Critical CVE-2026-21589 File Access Flaw in Data Center Products TechNadu · 2026-10-06

Researched and written by the R3KONX analysis desk from the cited primary material. Version numbers, scoring and mitigation detail are taken from Atlassian's advisory and verified against five independent accounts of it; one account's fixed-version strings were run together without separators and are read here against the advisory. No exploitation detail or proof of concept is reproduced. No exploitation was reported at the time of writing, and no credible internet-exposure count has been published, so none is given. Corrections to event@r3konx.asia.

R3KONX 2027

Work in this domain? So does the programme

4–6 May 2027, World Trade Centre Kuala Lumpur. Curated talks, hands-on training and The BattleGrid.