CVE-2026-35397: Jupyter path traversal leaks sibling

Q: Is CVE-2026-35397 actively exploited?

No confirmed active exploitation of CVE-2026-35397 has been reported, but organizations should still patch proactively.

Q: How to fix CVE-2026-35397?

1. Patch: Upgrade jupyter-server to ≥2.18.0 immediately — this is the only full fix. 2. Workaround (if patching is delayed): Rename all user root directories to use non-overlapping names with no shared prefixes — UUIDs or randomly generated identifiers eliminate the prefix-matching bypass entirely. 3. Audit: Search access logs for HTTP requests to /api/contents paths containing %2e%2e, %2F, or other URL-encoded path separator sequences; these indicate active exploitation attempts. 4. Network isolation: Restrict Jupyter Server to authenticated internal networks; never expose it directly to the public internet. 5. Inventory: Identify all internal deployments of jupyter-server and its dependent packages (1,862 downstream packages) using your SCA tooling and prioritize patching by multi-tenancy exposure.

Q: What systems are affected by CVE-2026-35397?

This vulnerability affects the following AI/ML architecture patterns: Multi-tenant Jupyter environments, ML development platforms, Training pipelines, Data science workspaces, Model development notebooks.

Q: What is the CVSS score for CVE-2026-35397?

CVE-2026-35397 has a CVSS v3.1 base score of 7.1 (HIGH). The EPSS exploitation probability is 0.54%.

CISO Take

Jupyter Server ≤2.17.0 contains a path traversal flaw (CWE-22) where the directory confinement logic uses prefix matching instead of exact matching — allowing any authenticated user to read, write, and delete files outside their designated root directory simply by crafting a URL-encoded `%2e%2e` request to the checkpoints API. With 1,862 downstream dependents across the ML/data science ecosystem and a working curl-based PoC already public, exposure is immediate and broad; multi-tenant JupyterHub deployments with predictable user directory naming (user1, user2…user10) are at highest risk of cross-tenant lateral access to model artifacts, training data, and notebooks containing credentials. Despite a low raw EPSS of 0.00036, this sits in the top 89th percentile for exploitation likelihood, and the trivial PoC requires zero AI/ML knowledge to execute. Patch to jupyter-server 2.18.0 now; if patching is delayed, rename user root directories to eliminate shared prefixes.

Sources: NVD GitHub Advisory EPSS ATLAS OpenSSF

What is the risk?

HIGH risk for any multi-tenant Jupyter deployment. Authentication is required (PR:L) and network complexity is rated high (AC:H), but the published PoC trivializes exploitation to a single curl command once the attacker has valid credentials. The impact is severe in shared ML platforms: confidentiality and integrity are both rated HIGH in the CVSS vector, and the prefix-bypass logic means a single-character username could grant access to a large number of sibling directories. Single-user local instances carry minimal risk. Risk is compounded by 9 prior CVEs in the same package and an OpenSSF Scorecard of only 4.8/10, indicating ongoing security hygiene concerns in the project.

How does the attack unfold?

Initial Access

Attacker authenticates to a shared Jupyter Server instance using legitimate user credentials on a multi-tenant ML platform.

AML.T0012

Reconnaissance

Attacker identifies predictable directory naming conventions used by the platform (e.g., user1…user10) and determines which sibling directories share a prefix with their own root.

AML.T0006

Exploitation

Attacker sends a crafted POST to /api/contents/%2e%2e/<sibling_dir>/<target>/checkpoints, exploiting the prefix-matching bug to escape root directory confinement.

AML.T0049

Data Exfiltration

Attacker reads, writes, or deletes files from sibling users' directories — extracting model weights, training data, credentials, and proprietary notebooks via the Jupyter REST API.

AML.T0025

Initial Access

Attacker authenticates to a shared Jupyter Server instance using legitimate user credentials on a multi-tenant ML platform.

AML.T0012

Reconnaissance

Attacker identifies predictable directory naming conventions used by the platform (e.g., user1…user10) and determines which sibling directories share a prefix with their own root.

AML.T0006

Exploitation

Attacker sends a crafted POST to /api/contents/%2e%2e/<sibling_dir>/<target>/checkpoints, exploiting the prefix-matching bug to escape root directory confinement.

AML.T0049

Data Exfiltration

Attacker reads, writes, or deletes files from sibling users' directories — extracting model weights, training data, credentials, and proprietary notebooks via the Jupyter REST API.

AML.T0025

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Jupyter	pip	—	No patch
13.2K OpenSSF 5.8 1.9K dependents Pushed 7d ago 79% patched ~9d to patch Full package profile →
Jupyter	pip	<= 2.17.0	`2.18.0`
13.2K OpenSSF 5.8 1.9K dependents Pushed 7d ago 79% patched ~9d to patch Full package profile →
Jupyter Notebook	pip	—	No patch
13.2K OpenSSF 5.8 3.0K dependents Pushed 7d ago 58% patched ~371d to patch Full package profile →

How severe is it?

CVSS 3.1

7.1 / 10

EPSS

0.5%

chance of exploitation in 30 days

Higher than 41% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AV Network

AC High

PR Low

UI None

S Unchanged

C High

I High

A Low

What should I do?

5 steps

Patch: Upgrade jupyter-server to ≥2.18.0 immediately — this is the only full fix.
Workaround (if patching is delayed): Rename all user root directories to use non-overlapping names with no shared prefixes — UUIDs or randomly generated identifiers eliminate the prefix-matching bypass entirely.
Audit: Search access logs for HTTP requests to /api/contents paths containing %2e%2e, %2F, or other URL-encoded path separator sequences; these indicate active exploitation attempts.
Network isolation: Restrict Jupyter Server to authenticated internal networks; never expose it directly to the public internet.
Inventory: Identify all internal deployments of jupyter-server and its dependent packages (1,862 downstream packages) using your SCA tooling and prioritize patching by multi-tenancy exposure.

What does CISA's SSVC say?

Decision Attend

Exploitation poc

Automatable No

Technical Impact total

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Data Extraction Privacy Violation Framework AML.T0025 - Exfiltration via Cyber Means AML.T0035 - AI Artifact Collection AML.T0037 - Data from Local System AML.T0049 - Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art. 9 - Risk management system

ISO 42001

A.6.2.5 - AI system operation — access control A.9.3 - Data acquisition and data quality — protection of training data

NIST AI RMF

GOVERN 1.2 - Accountability and transparency for AI risk management MANAGE 2.4 - Mechanisms for detecting and responding to AI risks

Frequently Asked Questions

What is CVE-2026-35397?

Jupyter Server ≤2.17.0 contains a path traversal flaw (CWE-22) where the directory confinement logic uses prefix matching instead of exact matching — allowing any authenticated user to read, write, and delete files outside their designated root directory simply by crafting a URL-encoded `%2e%2e` request to the checkpoints API. With 1,862 downstream dependents across the ML/data science ecosystem and a working curl-based PoC already public, exposure is immediate and broad; multi-tenant JupyterHub deployments with predictable user directory naming (user1, user2…user10) are at highest risk of cross-tenant lateral access to model artifacts, training data, and notebooks containing credentials. Despite a low raw EPSS of 0.00036, this sits in the top 89th percentile for exploitation likelihood, and the trivial PoC requires zero AI/ML knowledge to execute. Patch to jupyter-server 2.18.0 now; if patching is delayed, rename user root directories to eliminate shared prefixes.

Is CVE-2026-35397 actively exploited?

No confirmed active exploitation of CVE-2026-35397 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-35397?

1. Patch: Upgrade jupyter-server to ≥2.18.0 immediately — this is the only full fix. 2. Workaround (if patching is delayed): Rename all user root directories to use non-overlapping names with no shared prefixes — UUIDs or randomly generated identifiers eliminate the prefix-matching bypass entirely. 3. Audit: Search access logs for HTTP requests to /api/contents paths containing %2e%2e, %2F, or other URL-encoded path separator sequences; these indicate active exploitation attempts. 4. Network isolation: Restrict Jupyter Server to authenticated internal networks; never expose it directly to the public internet. 5. Inventory: Identify all internal deployments of jupyter-server and its dependent packages (1,862 downstream packages) using your SCA tooling and prioritize patching by multi-tenancy exposure.

What systems are affected by CVE-2026-35397?

This vulnerability affects the following AI/ML architecture patterns: Multi-tenant Jupyter environments, ML development platforms, Training pipelines, Data science workspaces, Model development notebooks.

What is the CVSS score for CVE-2026-35397?

CVE-2026-35397 has a CVSS v3.1 base score of 7.1 (HIGH). The EPSS exploitation probability is 0.54%.

What is the AI security impact?

Affected AI Architectures

Multi-tenant Jupyter environmentsML development platformsTraining pipelinesData science workspacesModel development notebooks

MITRE ATLAS Techniques

AML.T0025 Exfiltration via Cyber Means

AML.T0035 AI Artifact Collection

AML.T0037 Data from Local System

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Art. 9

ISO 42001: A.6.2.5, A.9.3

NIST AI RMF: GOVERN 1.2, MANAGE 2.4

What are the technical details?

Original Advisory

### Summary Jupyter Server <=2.17.0 can access directories sibling to the root directory, if it starts with the root dir's name. ### PoC Minimal: ``` . ├── test/ <- root directory. │ └── test.txt └── testtest/ └── secret.txt <- file to exfiltrate that we should not be able to access via API ``` ```bash HOST="http://localhost:8888" TOKEN="" SIBLING="testtest" TARGET="secret.txt" curl -s -X POST \ "$HOST/api/contents/%2e%2e/$SIBLING/$TARGET/checkpoints" \ -H "Authorization: token $TOKEN" ``` Full PoC by @stef41: https://gist.github.com/Yann-P/66d4982a965dee8fcb8dd89db29e7006 ### Impact It is possible for an authenticated user to access content outside the server's `root_dir` in siblings directories sharing the same prefix as the `root_dir`. The attacker can escalate access, reading, writing, and deleting from sibling directories. This can have a tangible impact for deployments using predictable naming scheme with multi-tenant server, for example `user1`, `user2`, `user3`, ..., `user10` etc, as `user1` could access and modify files of all `user10` - `user19` and higher. In a hypothetical system where users can choose a name of their folder, an attacker could choose a single-letter username to gain access to a significant number of sibling directories. ### Workarounds Use folder names that do not overlap. ### Acknowledgments Thank you to @stef41 for providing a useful PoC.

Exploitation Scenario

A junior data scientist with legitimate access to a shared JupyterHub ML platform discovers the platform names user home directories as ds-alice, ds-alice-research, ds-alice-dev. Knowing this, they issue a POST to `/api/contents/%2e%2e/ds-alice-research/proprietary_model.pkl/checkpoints` using their valid session token — the server resolves the path to the sibling directory and returns the file. Over the next hour, they enumerate and download model weights trained on proprietary datasets, notebooks containing AWS API keys used for S3 data access, and configuration files revealing internal infrastructure details. Using the write capability, they plant a backdoored version of a shared utility notebook, ensuring the payload executes the next time the victim opens their workspace.

Weaknesses (CWE)

CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') Primary

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.

[Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
[Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.