CVE-2026-8756: path traversal exposes ML training

CISO Take

CVE-2026-8756 is an unauthenticated path traversal in the Gradio preprocessing interface of fishaudio Bert-VITS2, a widely-used open-source voice synthesis model toolkit. The `generate_config` function accepts user-controlled `data_dir` input without sanitization, allowing remote attackers to traverse outside the intended working directory and read or write arbitrary files on the host. While EPSS data is not yet available and the vulnerability is absent from CISA KEV, the CVSS 7.3 HIGH score reflects no authentication required, no user interaction, and trivial network-accessible exploitation — a public proof-of-concept exploit has already been disclosed on GitHub, meaning the window before widespread scanning begins is short. Teams running Bert-VITS2 training or preprocessing pipelines should immediately restrict the Gradio interface to localhost-only binding (`server_name='127.0.0.1'`) and audit whether any instance is currently exposed on internal or external networks.

Sources: NVD ATLAS GitHub Advisory

What is the risk?

HIGH operational risk for ML engineering teams running Bert-VITS2 preprocessing pipelines. No authentication bypass needed — the Gradio interface accepts unauthenticated requests by default when server binding is not restricted. Combining network-accessible path traversal (CVSS AV:N/AC:L/PR:N/UI:N) with a live public PoC yields a near-certain exploitation scenario for any exposed instance. Training environments often store API keys, SSH keys, and cloud storage credentials in nearby configuration files, elevating the blast radius beyond the ML artifact scope. The absence of versioning in this project means no patched release exists; mitigations are configuration-only.

How does the attack unfold?

Discovery

Attacker scans the internet or internal network for Gradio interfaces on default port 7860 and identifies an exposed Bert-VITS2 preprocessing endpoint with no authentication.

AML.T0006

Exploitation

Using the public GitHub PoC, attacker submits a crafted request to the `generate_config` Gradio function with path traversal sequences in the `data_dir` argument, bypassing directory restrictions.

AML.T0049

Collection

Attacker reads arbitrary files from the host filesystem reachable by the process user — targeting `.env` files, Hugging Face tokens, SSH keys, and training dataset manifests.

AML.T0037

Exfiltration

Harvested credentials and AI artifacts are returned directly via HTTP responses from the Gradio interface, enabling follow-on access to model registries, cloud storage buckets, or compute infrastructure.

AML.T0025

Discovery

Attacker scans the internet or internal network for Gradio interfaces on default port 7860 and identifies an exposed Bert-VITS2 preprocessing endpoint with no authentication.

AML.T0006

Exploitation

Using the public GitHub PoC, attacker submits a crafted request to the `generate_config` Gradio function with path traversal sequences in the `data_dir` argument, bypassing directory restrictions.

AML.T0049

Collection

Attacker reads arbitrary files from the host filesystem reachable by the process user — targeting `.env` files, Hugging Face tokens, SSH keys, and training dataset manifests.

AML.T0037

Exfiltration

Harvested credentials and AI artifacts are returned directly via HTTP responses from the Gradio interface, enabling follow-on access to model registries, cloud storage buckets, or compute infrastructure.

AML.T0025

How severe is it?

CVSS 3.1

7.3 / 10

EPSS

0.5%

chance of exploitation in 30 days

Higher than 40% 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 Low

PR None

UI None

S Unchanged

C Low

I Low

A Low

What should I do?

6 steps

Immediately bind the Gradio interface to localhost: add server_name='127.0.0.1' to any gr.Blocks().launch() call in webui_preprocess.py.
Since the project has no versioning, pin to a commit after the vendor addresses this — track the repository for a fix commit post-disclosure.
Audit all running Bert-VITS2 instances using ss -tlnp | grep 7860 (default Gradio port) to identify exposed services.
Apply network-level controls: block inbound access to Gradio default ports (7860-7870) at the firewall or security group level.
For detection, monitor file access logs for path traversal patterns (../) in web server request logs or using auditd rules on sensitive directories.
Rotate any credentials stored in files accessible from the Bert-VITS2 working directory.

What does CISA's SSVC say?

Decision Track*

Exploitation poc

Automatable Yes

Technical Impact partial

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

How is it classified?

Data Extraction Code Execution Framework Training Data 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

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2.6 - Security of AI systems

NIST AI RMF

MANAGE 2.2 - Mechanisms are in place and applied to sustain the value of deployed AI systems

OWASP LLM Top 10

LLM06:2023 - Sensitive Information Disclosure

Frequently Asked Questions

What is CVE-2026-8756?

CVE-2026-8756 is an unauthenticated path traversal in the Gradio preprocessing interface of fishaudio Bert-VITS2, a widely-used open-source voice synthesis model toolkit. The `generate_config` function accepts user-controlled `data_dir` input without sanitization, allowing remote attackers to traverse outside the intended working directory and read or write arbitrary files on the host. While EPSS data is not yet available and the vulnerability is absent from CISA KEV, the CVSS 7.3 HIGH score reflects no authentication required, no user interaction, and trivial network-accessible exploitation — a public proof-of-concept exploit has already been disclosed on GitHub, meaning the window before widespread scanning begins is short. Teams running Bert-VITS2 training or preprocessing pipelines should immediately restrict the Gradio interface to localhost-only binding (`server_name='127.0.0.1'`) and audit whether any instance is currently exposed on internal or external networks.

Is CVE-2026-8756 actively exploited?

No confirmed active exploitation of CVE-2026-8756 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-8756?

1. Immediately bind the Gradio interface to localhost: add `server_name='127.0.0.1'` to any `gr.Blocks().launch()` call in `webui_preprocess.py`. 2. Since the project has no versioning, pin to a commit after the vendor addresses this — track the repository for a fix commit post-disclosure. 3. Audit all running Bert-VITS2 instances using `ss -tlnp | grep 7860` (default Gradio port) to identify exposed services. 4. Apply network-level controls: block inbound access to Gradio default ports (7860-7870) at the firewall or security group level. 5. For detection, monitor file access logs for path traversal patterns (`../`) in web server request logs or using auditd rules on sensitive directories. 6. Rotate any credentials stored in files accessible from the Bert-VITS2 working directory.

What systems are affected by CVE-2026-8756?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, ml ui interfaces, model serving.

What is the CVSS score for CVE-2026-8756?

CVE-2026-8756 has a CVSS v3.1 base score of 7.3 (HIGH). The EPSS exploitation probability is 0.51%.

What is the AI security impact?

Affected AI Architectures

training pipelinesml ui interfacesmodel serving

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: Article 15

ISO 42001: A.6.2.6

NIST AI RMF: MANAGE 2.2

OWASP LLM Top 10: LLM06:2023

What are the technical details?

Original Advisory

A vulnerability has been found in fishaudio Bert-VITS2 up to 8f7fbd8c4770965225d258db548da27dc8dd934c. The impacted element is the function generate_config of the file webui_preprocess.py of the component Gradio Interface. Such manipulation of the argument data_dir leads to path traversal. The attack can be launched remotely. The exploit has been disclosed to the public and may be used. This product does not use versioning. This is why information about affected and unaffected releases are unavailable. The vendor was contacted early about this disclosure but did not respond in any way.

Exploitation Scenario

An attacker scans for internet-exposed services on port 7860 (Gradio default) and identifies a Bert-VITS2 preprocessing instance with no authentication. Using the public GitHub PoC, they submit a crafted POST request to the `generate_config` Gradio endpoint with `data_dir` set to a path traversal sequence such as `../../../../../../home/user/.ssh/` or `../../opt/ml/.env`. The server processes the traversal without validation and returns file contents or writes a malicious config. The attacker exfiltrates Hugging Face API tokens, cloud storage credentials, or SSH private keys found in adjacent directories — enabling persistent access to model registries, training data buckets, or the underlying compute infrastructure.

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.