CVE-2026-113743: tt-metal: path traversal in checkpoint loader

MEDIUM

Published October 21, 2026

CISO Take

Tenstorrent tt-metal, the runtime for Tenstorrent AI accelerators, has a path traversal flaw in load_hf_state_dict_filtered, which trusts the weight_map in a Hugging Face checkpoint index file without validating the paths in it. A crafted checkpoint can make the loader read files outside the intended model directory. This is a medium-severity issue (CVSS 6.3) that needs user interaction, in practice loading an untrusted model. It is not in CISA KEV and no EPSS score is available. A public write-up of the exploit exists on GitHub, so the barrier to attack is low, although no Nuclei template or SSVC decision is recorded. The tracked package risk is low (0/100) and no other CVEs are recorded for it. The exposure is concentrated in teams that run Tenstorrent hardware and pull third-party models. Upgrade to v0.81.0-dev20261016 or later (patch commit 4851ccf) and only load checkpoints from trusted, verified sources until then.

Sources:NVDATLASvuldb.comGitHub (tenstorrent/tt-metal)

What is the risk?

Medium. The attack is network-reachable with low complexity and no privileges, but it needs a victim to load a crafted checkpoint (UI:R). Impact is rated low across confidentiality, integrity and availability. A public exploit write-up raises the likelihood of opportunistic use, but there is no KEV listing, no scanner template and no EPSS data to suggest active exploitation. Exposure is limited to deployments that run tt-metal's tt_transformers models with externally sourced checkpoints, which is a narrow population.

How does the attack unfold?

Artifact staging

Attacker publishes a model repository with a crafted weight_map containing path traversal sequences.

AML.T0010.003

User execution

A victim or automated pipeline downloads the model and loads it with tt-metal's tt_transformers.

AML.T0011.000

Exploitation

load_hf_state_dict_filtered resolves weight_map paths outside the model directory without validation.

Impact

Files outside the model directory are accessed, which can expose local data or disrupt model loading.

AML.T0037

Artifact staging

Attacker publishes a model repository with a crafted weight_map containing path traversal sequences.

AML.T0010.003

User execution

A victim or automated pipeline downloads the model and loads it with tt-metal's tt_transformers.

AML.T0011.000

Exploitation

load_hf_state_dict_filtered resolves weight_map paths outside the model directory without validation.

Impact

Files outside the model directory are accessed, which can expose local data or disrupt model loading.

AML.T0037

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
tt-metal	—	—	No patch

Do you use tt-metal? You're affected.

How severe is it?

CVSS 3.1

6.3 / 10

EPSS

0.4%

chance of exploitation in 30 days

Higher than 39% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What is the attack surface?

AVNetwork

ACLow

PRNone

UIRequired

SUnchanged

CLow

ILow

ALow

What should I do?

1 step

1) Upgrade tt-metal to v0.81.0-dev20261016 or later, or apply commit 4851ccf347f6903c3d819eed60705600386e52c3 (PR 60634). 2) Until patched, load checkpoints only from trusted, pinned and verified sources, and review model.safetensors.index.json for absolute paths or '..' segments in weight_map values. 3) Run model-loading jobs as an unprivileged user in a container or sandbox with a restricted filesystem view and no access to secrets or credentials. 4) Inventory hosts running tt-metal versions up to 0.80.0. 5) Monitor for file access outside model directories during checkpoint loads, and for unexpected weight index files in model downloads.

How is it classified?

Supply ChainData LeakageFrameworkModelInferenceAML.T0010.003- Model AML.T0011.000- Unsafe AI Artifacts AML.T0037- Data from Local System

Which compliance frameworks are affected?

This CVE is relevant to:

ISO 42001

A.10.3-Suppliers

NIST AI RMF

MANAGE 2.3-Responding to unknown risks and third-party component risks

OWASP LLM Top 10

LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-113743?

Is CVE-2026-113743 actively exploited?

No confirmed active exploitation of CVE-2026-113743 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-113743?

1) Upgrade tt-metal to v0.81.0-dev20261016 or later, or apply commit 4851ccf347f6903c3d819eed60705600386e52c3 (PR 60634). 2) Until patched, load checkpoints only from trusted, pinned and verified sources, and review model.safetensors.index.json for absolute paths or '..' segments in weight_map values. 3) Run model-loading jobs as an unprivileged user in a container or sandbox with a restricted filesystem view and no access to secrets or credentials. 4) Inventory hosts running tt-metal versions up to 0.80.0. 5) Monitor for file access outside model directories during checkpoint loads, and for unexpected weight index files in model downloads.

What systems are affected by CVE-2026-113743?

This vulnerability affects the following AI/ML architecture patterns: model serving, inference on Tenstorrent accelerators, model loading pipelines, development and CI environments.

What is the CVSS score for CVE-2026-113743?

CVE-2026-113743 has a CVSS v3.1 base score of 6.3 (MEDIUM). The EPSS exploitation probability is 0.44%.

What is the AI security impact?

Affected AI Architectures

model servinginference on Tenstorrent acceleratorsmodel loading pipelinesdevelopment and CI environments

MITRE ATLAS Techniques

AML.T0010.003Model

AML.T0011.000Unsafe AI Artifacts

AML.T0037Data from Local System

Compliance Controls Affected

ISO 42001: A.10.3

NIST AI RMF: MANAGE 2.3

OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

A weakness has been identified in Tenstorrent tt-metal up to 0.80.0. The affected element is the function load_hf_state_dict_filtered of the file models/tt_transformers/tt/load_checkpoints.py. This manipulation of the argument weight_map causes path traversal. The attack may be initiated remotely. The exploit has been made available to the public and could be used for attacks. Upgrading to version 0.81.0-dev20261016 is sufficient to fix this issue. Patch name: 4851ccf347f6903c3d819eed60705600386e52c3. It is suggested to upgrade the affected component.

Exploitation Scenario

An attacker publishes a model repository that looks legitimate on a public hub, with a modified index file whose weight_map entries contain traversal sequences such as ../../ paths. An engineer or automated pipeline running tt-metal on Tenstorrent hardware downloads and loads the model. The vulnerable loader resolves those paths outside the checkpoint directory and reads unintended files on the host. This can disrupt model loading, corrupt loaded state or expose local data, depending on the handling of the resolved files.

Weaknesses (CWE)

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

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.