Hugging Face
Monthly
Loading a malicious model into Unsloth Zoo 2025.9.9 through 2026.8.13 (as shipped in Unsloth 2025.9.9 through 2026.8.19) lets that model's config.json run arbitrary Python in the loading user's process, because a nested model_type value containing a newline escapes into an exec()-generated import statement. Exploitation requires the victim to load an attacker-controlled model for training or inference - a user-interaction gate, but no authentication on the attacker side (CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H; vendor score 8.6). No public exploit code has been identified at time of analysis, and the practical exposure is concentrated in pipelines that automatically ingest third-party or user-supplied models from public hubs rather than those loading only vetted first-party checkpoints.
Arbitrary Python code execution in LMDeploy 0.12.1 through 0.12.2 lets an attacker who publishes a malicious model on HuggingFace Hub run code on any machine that loads it, because the model's `quantization_config.quant_dtype` string is interpolated into `eval(f'torch.{quant_dtype}')` at lmdeploy/pytorch/config.py:620 without any validation. The CVSS vector (AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H, 8.8) reflects that exploitation is remote and unauthenticated but requires the victim to actively load the attacker-supplied model, confirmed by our independent assessment that UI:R is the limiting factor and that the `_update_torch_dtype` whitelist at line 53 only guards `torch_dtype`, not `quant_dtype`. Publicly available proof-of-concept code exists in the vendor advisory (repo-level proof using a mocked HF config that yields `uid=0(root)`), but there is no confirmed actively exploited (CISA KEV) status, and no EPSS score was supplied; a fixed release, 0.12.3, is available.
Out-of-bounds write in Hugging Face tokenizers v0.23.1 allows any actor who can supply a crafted tokenizer.json file to abort the loading process immediately, causing denial of service in Rust applications and FFI embeddings before any inference occurs. The defect in BpeBuilder::build (tokenizers/src/models/bpe/model.rs) allocates a scratch buffer sized to the longest vocabulary key, then writes concatenated BPE merge pairs into it without bounding the write to the buffer size; a merge whose result exceeds that length triggers a Rust panic and process abort (CWE-787). A secondary usize underflow at the same code site, activated when continuing_subword_prefix is set and a merge token is shorter than the prefix, produces a panic in debug builds and potential memory corruption in release builds, elevating the secondary impact beyond availability alone. No public exploit code or CISA KEV listing has been identified at time of analysis; a fix appears in the v0.23.2 source.
Unsafe pickle deserialization in Hugging Face PEFT's LoRA-GA and CorDA initialization modules allows arbitrary code execution when a victim loads a malicious cache or covariance file. The affected code paths in corda.py (lines ~102 and ~163) and loraga.py (line ~101) call torch.load() without weights_only=True, bypassing the safe-loading wrapper that PEFT enforces elsewhere in the codebase. An attacker who can supply or substitute a pre-computed LoRA-GA or CorDA cache file - for example via a shared model repository or poisoned download - achieves full code execution in the victim's ML environment upon initialization. No public exploit code or active exploitation (CISA KEV) has been identified at time of analysis.
Arbitrary file write via path traversal in HuggingFace Transformers (<=5.8.0.dev0, fixed in 5.10.0) enables a supply-chain attacker to overwrite arbitrary files on a victim's filesystem. The flaw exists in `save_pretrained()` methods of `PreTrainedTokenizerBase` and `ProcessorMixin`, where dictionary keys from the `chat_template` field in a downloaded `tokenizer_config.json` are used verbatim as filenames with no path sanitization. An attacker need only publish a malicious model repository on HuggingFace Hub; exploitation is triggered when the victim calls `save_pretrained()` after loading the poisoned model. No active exploitation is confirmed (not in CISA KEV), but the attack surface spans all practitioners using the affected `save_pretrained()` path across multiple popular processor families.
Remote code execution in Hugging Face Transformers 5.2.0 allows a malicious model repository to bypass the user's explicit trust_remote_code=False safeguard when loading a LightGlue model via AutoModel.from_pretrained(). The LightGlueConfig deserializes the trust_remote_code flag from the untrusted config.json and propagates the attacker-controlled value into a nested AutoConfig.from_pretrained() call, enabling execution of arbitrary attacker-supplied Python during model initialization. Rated CVSS 9.6 (AV:N/AC:L/PR:N/UI:R) with publicly available exploit code exists via the Huntr disclosure, though EPSS is currently 0.07% (22th percentile) and the CVE is not on CISA KEV.
Remote code execution in HuggingFace Transformers prior to 5.3.0 allows attackers to achieve arbitrary code execution on a victim's machine by publishing a malicious model whose config.json sets the `_attn_implementation_internal` field to an attacker-controlled Hub repository. When the victim calls the standard `AutoModelForCausalLM.from_pretrained()` API, the library silently downloads and executes Python kernels from that repository with the victim's privileges, bypassing the `trust_remote_code` safety gate. No public exploit is identified at time of analysis (EPSS 0.03%, SSVC exploitation: none), but the technical impact is total and the attack uses the documented, default usage pattern.
Remote code execution in HuggingFace Diffusers library (versions < 0.38.0) allows attackers to execute arbitrary Python code when victims load malicious pipelines from Hugging Face Hub repositories. The vulnerability bypasses the trust_remote_code=True safeguard through a type coercion flaw where None values are interpolated as 'None.py' filenames. Attackers can achieve silent code execution by publishing repositories containing a malicious None.py file alongside legitimate-looking configuration, requiring only that victims call DiffusionPipeline.from_pretrained() on the attacker's repository. EPSS data not available; no public exploit identified at time of analysis. Vendor-released patch: version 0.38.0.
Unauthenticated remote code execution in HuggingFace LeRobot (versions 0 through 0.5.1) stems from pickle.loads() being called on attacker-controlled bytes received over the async inference gRPC pipeline, which runs without authentication or TLS. Any network-reachable attacker can execute arbitrary code on either the policy server or the robot client by delivering a malicious pickle payload through the SendPolicyInstructions, SendObservations, or GetActions RPCs. Publicly available exploit code exists (chocapikk write-up) and SSVC rates technical impact as total and the flaw as automatable, though EPSS remains very low (0.06%) reflecting limited real-world telemetry rather than low technical severity.
Remote code execution in HuggingFace Transformers library allows arbitrary code execution via malicious checkpoint files. The `_load_rng_state()` method in the `Trainer` class calls `torch.load()` without the `weights_only=True` parameter, enabling deserialization attacks when PyTorch versions below 2.6 are used with torch>=2.2. An attacker can craft a malicious `rng_state.pth` checkpoint file that executes arbitrary code when loaded by an application using affected Transformers versions. The fix is available in version v5.0.0rc3, and no public exploit has been independently confirmed at time of analysis.
Arbitrary file read in Hugging Face Transformers allows an attacker who controls a model checkpoint index file to exfiltrate files from the host filesystem outside the intended model directory. The library fails to sanitize shard filenames in checkpoint index files before joining them to directory paths, meaning maliciously crafted entries containing parent-directory sequences or absolute paths resolve to unintended locations. Any user or automated pipeline that loads an attacker-supplied checkpoint - for example, by downloading a model from an untrusted Hub repository - can have host files disclosed without code execution. No public exploit has been identified at time of analysis, and the vulnerability is not listed in CISA KEV.
Path traversal in Hugging Face Accelerate through 1.14.0 exposes two distinct attack outcomes when a user loads a crafted sharded checkpoint: sensitive file disclosure via relative (../) or absolute path entries in the weight_map index, and indefinite process blocking via named pipe references enabling denial of service. A publicly available proof-of-concept exists at GitHub issue #4067, reported by VulnCheck. With no system-level privileges required (PR:N per CVSS 4.0), this is a supply chain risk for any ML practitioner who loads untrusted model checkpoints - particularly from public model registries - without needing any foothold on the target system beyond delivering the malicious checkpoint.
Path traversal in Hugging Face Datasets up to 5.0.0 allows attackers to read arbitrary local files when a victim processes a crafted dataset. A publicly available exploit exists (CVE-2026-66007), though EPSS exploitation probability is low at 0.52% and user interaction is required. Patched in commit f989ef9.
Symlink following in HuggingFace Datasets through version 5.00 allows local low-privileged attackers to redirect archive extraction to arbitrary filesystem locations by pre-planting symlinks at predictable output paths in the shared cache directory. When a higher-privileged user or automated service subsequently triggers Extractor.extract(), the library writes archive contents through the attacker-controlled symlink, enabling overwrite of sensitive system files and potential privilege escalation or code execution. A publicly available exploit exists and an upstream fix is available via commit ad2d853, though no confirmed tagged release version has been independently verified.
Path traversal in HuggingFace Diffusers through 0.39.0 enables arbitrary file reads when a user or pipeline loads a maliciously crafted model. The `_get_checkpoint_shard_files` function in `hub_utils.py` blindly resolves `weight_map` shard filenames from model index JSON without validating for path separators or absolute paths, allowing an attacker who controls a model repository to escape the model directory and read any file accessible to the Diffusers process. No active exploitation confirmed (not in CISA KEV), but a public POC issue exists on GitHub and the attack surface includes the widely-used HuggingFace Hub ecosystem.
Server-side request forgery in HuggingFace text-generation-inference through version 3.3.7 enables unauthenticated remote attackers to coerce the server into issuing arbitrary outbound HTTP GET requests via a crafted image_url value submitted to the OpenAI-compatible multimodal chat completions endpoint. The fetch_image function in router/src/validation.rs applies no destination validation, and the reqwest client's default redirect-following behavior allows attackers to chain redirects to bypass any scheme-level controls, reaching cloud instance-metadata services (e.g., AWS IMDSv1 at 169.254.169.254) to steal IAM credentials or enumerate internal port services. Publicly available exploit code exists (VulnCheck/geo-chen); no KEV listing at time of analysis, but the combination of unauthenticated access, cloud credential theft potential, and working POC makes this a genuine priority for any cloud-hosted TGI deployment.
Code injection in HuggingFace smolagents 1.25.0.dev0 allows remote attackers without authentication to execute arbitrary code through incomplete remediation of CVE-2025-9959 in the local Python executor component. The vulnerability affects the evaluate_augassign, evaluate_call, and evaluate_with functions in src/smolagents/local_python_executor.py, with publicly available exploit code and active public disclosure despite lack of vendor response.
Open redirect in Gradio's OAuth implementation allows unauthenticated attackers to redirect users to arbitrary external URLs through the unvalidated _target_url parameter on /logout and /login/callback endpoints in applications with OAuth enabled. This affects Gradio versions prior to 6.6.0 running on Hugging Face Spaces with gr.LoginButton, enabling phishing attacks or credential theft. The vulnerability has been patched in version 6.6.0 by sanitizing the parameter to only accept relative URLs.
Gradio is an open-source Python package designed for quick prototyping. Starting in version 4.16.0 and prior to version 6.6.0, Gradio applications running outside of Hugging Face Spaces automatically enable "mocked" OAuth routes when OAuth components (e.g.
Server-side request forgery in Hugging Face smolagents 1.24.0 allows authenticated attackers to manipulate the LocalPythonExecutor's requests.get/requests.post functions, enabling remote exploitation without user interaction. Public exploit code exists for this vulnerability, and no patch is currently available from the vendor despite early disclosure notification.
XPath injection in Hugging Face Smolagents 1.20.0 lets an attacker who can influence the text passed to the vision web browser's Ctrl-F style search (search_item_ctrl_f) rewrite the underlying XPath query, bypassing intended search filters and pulling in unintended DOM elements. The flaw is reachable without authentication (CVSS PR:N) but requires the agent or user to trigger the search with attacker-influenced text (UI:R), so it is not passively exploitable; impact is bounded to information disclosure from the queried page, manipulation of agent search results, and disruption of automation workflows rather than code execution. Publicly available exploit code exists, but there is no confirmed active exploitation (not in CISA KEV), and EPSS is low at 0.28% (19th percentile), making this a moderate, situational risk rather than a high-priority emergency.
The huggingface/transformers library, versions prior to 4.53.0, is vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically within the `normalize_numbers()` method of the `EnglishNormalizer`. Rated medium severity (CVSS 5.3), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically affecting the MarianTokenizer's `remove_language_code()` method. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability exists in the Hugging Face Transformers library, specifically in the `convert_tf_weight_name_to_pt_weight_name()` function. Rated medium severity (CVSS 5.3), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically within the DonutProcessor class's `token2json()` method. This vulnerability affects versions 4.50.3 and earlier, and is fixed in version 4.52.1. The issue arises from the regex pattern `<s_(.*?)>` which can be exploited to cause excessive CPU consumption through crafted input strings due to catastrophic backtracking. This vulnerability can lead to service disruption, resource exhaustion, and potential API service vulnerabilities, impacting document processing tasks using the Donut model.
Hugging Face Transformers versions up to 4.49.0 are affected by an improper input validation vulnerability in the `image_utils.py` file. The vulnerability arises from insecure URL validation using the `startswith()` method, which can be bypassed through URL username injection. This allows attackers to craft URLs that appear to be from YouTube but resolve to malicious domains, potentially leading to phishing attacks, malware distribution, or data exfiltration. The issue is fixed in version 4.52.1.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically in the `get_imports()` function within `dynamic_module_utils.py`. This vulnerability affects versions 4.49.0 and is fixed in version 4.51.0. The issue arises from a regular expression pattern `\s*try\s*:.*?except.*?:` used to filter out try/except blocks from Python code, which can be exploited to cause excessive CPU consumption through crafted input strings due to catastrophic backtracking. This vulnerability can lead to remote code loading disruption, resource exhaustion in model serving, supply chain attack vectors, and development pipeline disruption.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically in the `get_configuration_file()` function within the `transformers.configuration_utils` module. The affected version is 4.49.0, and the issue is resolved in version 4.51.0. The vulnerability arises from the use of a regular expression pattern `config\.(.*)\.json` that can be exploited to cause excessive CPU consumption through crafted input strings, leading to catastrophic backtracking. This can result in model serving disruption, resource exhaustion, and increased latency in applications using the library.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the huggingface/transformers repository, specifically in version 4.49.0. The vulnerability is due to inefficient regular expression complexity in the `SETTING_RE` variable within the `transformers/commands/chat.py` file. The regex contains repetition groups and non-optimized quantifiers, leading to exponential backtracking when processing 'almost matching' payloads. This can degrade application performance and potentially result in a denial-of-service (DoS) when handling specially crafted input strings. The issue is fixed in version 4.51.0.
A vulnerability in the `preprocess_string()` function of the `transformers.testing_utils` module in huggingface/transformers version v4.48.3 allows for a Regular Expression Denial of Service (ReDoS). Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file `tokenization_gpt_neox_japanese.py` of the. Rated medium severity (CVSS 6.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file tokenization_nougat_fast.py. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity.
Hugging Face Transformers Trax Model Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
Hugging Face Transformers MaskFormer Model Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
Hugging Face Transformers MobileViTV2 Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
The huggingface/transformers library is vulnerable to arbitrary code execution through deserialization of untrusted data within the `load_repo_checkpoint()` function of the `TFPreTrainedModel()`. Rated critical severity (CVSS 9.6), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
Deserialization of Untrusted Data in GitHub repository huggingface/transformers prior to 4.36. Rated high severity (CVSS 7.8), this vulnerability is no authentication required, low attack complexity. Public exploit code available.
Deserialization of Untrusted Data in GitHub repository huggingface/transformers prior to 4.36. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, low attack complexity. Public exploit code available.
Insecure Temporary File in GitHub repository huggingface/transformers prior to 4.30.0. Rated medium severity (CVSS 4.7). Public exploit code available.
Loading a malicious model into Unsloth Zoo 2025.9.9 through 2026.8.13 (as shipped in Unsloth 2025.9.9 through 2026.8.19) lets that model's config.json run arbitrary Python in the loading user's process, because a nested model_type value containing a newline escapes into an exec()-generated import statement. Exploitation requires the victim to load an attacker-controlled model for training or inference - a user-interaction gate, but no authentication on the attacker side (CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H; vendor score 8.6). No public exploit code has been identified at time of analysis, and the practical exposure is concentrated in pipelines that automatically ingest third-party or user-supplied models from public hubs rather than those loading only vetted first-party checkpoints.
Arbitrary Python code execution in LMDeploy 0.12.1 through 0.12.2 lets an attacker who publishes a malicious model on HuggingFace Hub run code on any machine that loads it, because the model's `quantization_config.quant_dtype` string is interpolated into `eval(f'torch.{quant_dtype}')` at lmdeploy/pytorch/config.py:620 without any validation. The CVSS vector (AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H, 8.8) reflects that exploitation is remote and unauthenticated but requires the victim to actively load the attacker-supplied model, confirmed by our independent assessment that UI:R is the limiting factor and that the `_update_torch_dtype` whitelist at line 53 only guards `torch_dtype`, not `quant_dtype`. Publicly available proof-of-concept code exists in the vendor advisory (repo-level proof using a mocked HF config that yields `uid=0(root)`), but there is no confirmed actively exploited (CISA KEV) status, and no EPSS score was supplied; a fixed release, 0.12.3, is available.
Out-of-bounds write in Hugging Face tokenizers v0.23.1 allows any actor who can supply a crafted tokenizer.json file to abort the loading process immediately, causing denial of service in Rust applications and FFI embeddings before any inference occurs. The defect in BpeBuilder::build (tokenizers/src/models/bpe/model.rs) allocates a scratch buffer sized to the longest vocabulary key, then writes concatenated BPE merge pairs into it without bounding the write to the buffer size; a merge whose result exceeds that length triggers a Rust panic and process abort (CWE-787). A secondary usize underflow at the same code site, activated when continuing_subword_prefix is set and a merge token is shorter than the prefix, produces a panic in debug builds and potential memory corruption in release builds, elevating the secondary impact beyond availability alone. No public exploit code or CISA KEV listing has been identified at time of analysis; a fix appears in the v0.23.2 source.
Unsafe pickle deserialization in Hugging Face PEFT's LoRA-GA and CorDA initialization modules allows arbitrary code execution when a victim loads a malicious cache or covariance file. The affected code paths in corda.py (lines ~102 and ~163) and loraga.py (line ~101) call torch.load() without weights_only=True, bypassing the safe-loading wrapper that PEFT enforces elsewhere in the codebase. An attacker who can supply or substitute a pre-computed LoRA-GA or CorDA cache file - for example via a shared model repository or poisoned download - achieves full code execution in the victim's ML environment upon initialization. No public exploit code or active exploitation (CISA KEV) has been identified at time of analysis.
Arbitrary file write via path traversal in HuggingFace Transformers (<=5.8.0.dev0, fixed in 5.10.0) enables a supply-chain attacker to overwrite arbitrary files on a victim's filesystem. The flaw exists in `save_pretrained()` methods of `PreTrainedTokenizerBase` and `ProcessorMixin`, where dictionary keys from the `chat_template` field in a downloaded `tokenizer_config.json` are used verbatim as filenames with no path sanitization. An attacker need only publish a malicious model repository on HuggingFace Hub; exploitation is triggered when the victim calls `save_pretrained()` after loading the poisoned model. No active exploitation is confirmed (not in CISA KEV), but the attack surface spans all practitioners using the affected `save_pretrained()` path across multiple popular processor families.
Remote code execution in Hugging Face Transformers 5.2.0 allows a malicious model repository to bypass the user's explicit trust_remote_code=False safeguard when loading a LightGlue model via AutoModel.from_pretrained(). The LightGlueConfig deserializes the trust_remote_code flag from the untrusted config.json and propagates the attacker-controlled value into a nested AutoConfig.from_pretrained() call, enabling execution of arbitrary attacker-supplied Python during model initialization. Rated CVSS 9.6 (AV:N/AC:L/PR:N/UI:R) with publicly available exploit code exists via the Huntr disclosure, though EPSS is currently 0.07% (22th percentile) and the CVE is not on CISA KEV.
Remote code execution in HuggingFace Transformers prior to 5.3.0 allows attackers to achieve arbitrary code execution on a victim's machine by publishing a malicious model whose config.json sets the `_attn_implementation_internal` field to an attacker-controlled Hub repository. When the victim calls the standard `AutoModelForCausalLM.from_pretrained()` API, the library silently downloads and executes Python kernels from that repository with the victim's privileges, bypassing the `trust_remote_code` safety gate. No public exploit is identified at time of analysis (EPSS 0.03%, SSVC exploitation: none), but the technical impact is total and the attack uses the documented, default usage pattern.
Remote code execution in HuggingFace Diffusers library (versions < 0.38.0) allows attackers to execute arbitrary Python code when victims load malicious pipelines from Hugging Face Hub repositories. The vulnerability bypasses the trust_remote_code=True safeguard through a type coercion flaw where None values are interpolated as 'None.py' filenames. Attackers can achieve silent code execution by publishing repositories containing a malicious None.py file alongside legitimate-looking configuration, requiring only that victims call DiffusionPipeline.from_pretrained() on the attacker's repository. EPSS data not available; no public exploit identified at time of analysis. Vendor-released patch: version 0.38.0.
Unauthenticated remote code execution in HuggingFace LeRobot (versions 0 through 0.5.1) stems from pickle.loads() being called on attacker-controlled bytes received over the async inference gRPC pipeline, which runs without authentication or TLS. Any network-reachable attacker can execute arbitrary code on either the policy server or the robot client by delivering a malicious pickle payload through the SendPolicyInstructions, SendObservations, or GetActions RPCs. Publicly available exploit code exists (chocapikk write-up) and SSVC rates technical impact as total and the flaw as automatable, though EPSS remains very low (0.06%) reflecting limited real-world telemetry rather than low technical severity.
Remote code execution in HuggingFace Transformers library allows arbitrary code execution via malicious checkpoint files. The `_load_rng_state()` method in the `Trainer` class calls `torch.load()` without the `weights_only=True` parameter, enabling deserialization attacks when PyTorch versions below 2.6 are used with torch>=2.2. An attacker can craft a malicious `rng_state.pth` checkpoint file that executes arbitrary code when loaded by an application using affected Transformers versions. The fix is available in version v5.0.0rc3, and no public exploit has been independently confirmed at time of analysis.
Arbitrary file read in Hugging Face Transformers allows an attacker who controls a model checkpoint index file to exfiltrate files from the host filesystem outside the intended model directory. The library fails to sanitize shard filenames in checkpoint index files before joining them to directory paths, meaning maliciously crafted entries containing parent-directory sequences or absolute paths resolve to unintended locations. Any user or automated pipeline that loads an attacker-supplied checkpoint - for example, by downloading a model from an untrusted Hub repository - can have host files disclosed without code execution. No public exploit has been identified at time of analysis, and the vulnerability is not listed in CISA KEV.
Path traversal in Hugging Face Accelerate through 1.14.0 exposes two distinct attack outcomes when a user loads a crafted sharded checkpoint: sensitive file disclosure via relative (../) or absolute path entries in the weight_map index, and indefinite process blocking via named pipe references enabling denial of service. A publicly available proof-of-concept exists at GitHub issue #4067, reported by VulnCheck. With no system-level privileges required (PR:N per CVSS 4.0), this is a supply chain risk for any ML practitioner who loads untrusted model checkpoints - particularly from public model registries - without needing any foothold on the target system beyond delivering the malicious checkpoint.
Path traversal in Hugging Face Datasets up to 5.0.0 allows attackers to read arbitrary local files when a victim processes a crafted dataset. A publicly available exploit exists (CVE-2026-66007), though EPSS exploitation probability is low at 0.52% and user interaction is required. Patched in commit f989ef9.
Symlink following in HuggingFace Datasets through version 5.00 allows local low-privileged attackers to redirect archive extraction to arbitrary filesystem locations by pre-planting symlinks at predictable output paths in the shared cache directory. When a higher-privileged user or automated service subsequently triggers Extractor.extract(), the library writes archive contents through the attacker-controlled symlink, enabling overwrite of sensitive system files and potential privilege escalation or code execution. A publicly available exploit exists and an upstream fix is available via commit ad2d853, though no confirmed tagged release version has been independently verified.
Path traversal in HuggingFace Diffusers through 0.39.0 enables arbitrary file reads when a user or pipeline loads a maliciously crafted model. The `_get_checkpoint_shard_files` function in `hub_utils.py` blindly resolves `weight_map` shard filenames from model index JSON without validating for path separators or absolute paths, allowing an attacker who controls a model repository to escape the model directory and read any file accessible to the Diffusers process. No active exploitation confirmed (not in CISA KEV), but a public POC issue exists on GitHub and the attack surface includes the widely-used HuggingFace Hub ecosystem.
Server-side request forgery in HuggingFace text-generation-inference through version 3.3.7 enables unauthenticated remote attackers to coerce the server into issuing arbitrary outbound HTTP GET requests via a crafted image_url value submitted to the OpenAI-compatible multimodal chat completions endpoint. The fetch_image function in router/src/validation.rs applies no destination validation, and the reqwest client's default redirect-following behavior allows attackers to chain redirects to bypass any scheme-level controls, reaching cloud instance-metadata services (e.g., AWS IMDSv1 at 169.254.169.254) to steal IAM credentials or enumerate internal port services. Publicly available exploit code exists (VulnCheck/geo-chen); no KEV listing at time of analysis, but the combination of unauthenticated access, cloud credential theft potential, and working POC makes this a genuine priority for any cloud-hosted TGI deployment.
Code injection in HuggingFace smolagents 1.25.0.dev0 allows remote attackers without authentication to execute arbitrary code through incomplete remediation of CVE-2025-9959 in the local Python executor component. The vulnerability affects the evaluate_augassign, evaluate_call, and evaluate_with functions in src/smolagents/local_python_executor.py, with publicly available exploit code and active public disclosure despite lack of vendor response.
Open redirect in Gradio's OAuth implementation allows unauthenticated attackers to redirect users to arbitrary external URLs through the unvalidated _target_url parameter on /logout and /login/callback endpoints in applications with OAuth enabled. This affects Gradio versions prior to 6.6.0 running on Hugging Face Spaces with gr.LoginButton, enabling phishing attacks or credential theft. The vulnerability has been patched in version 6.6.0 by sanitizing the parameter to only accept relative URLs.
Gradio is an open-source Python package designed for quick prototyping. Starting in version 4.16.0 and prior to version 6.6.0, Gradio applications running outside of Hugging Face Spaces automatically enable "mocked" OAuth routes when OAuth components (e.g.
Server-side request forgery in Hugging Face smolagents 1.24.0 allows authenticated attackers to manipulate the LocalPythonExecutor's requests.get/requests.post functions, enabling remote exploitation without user interaction. Public exploit code exists for this vulnerability, and no patch is currently available from the vendor despite early disclosure notification.
XPath injection in Hugging Face Smolagents 1.20.0 lets an attacker who can influence the text passed to the vision web browser's Ctrl-F style search (search_item_ctrl_f) rewrite the underlying XPath query, bypassing intended search filters and pulling in unintended DOM elements. The flaw is reachable without authentication (CVSS PR:N) but requires the agent or user to trigger the search with attacker-influenced text (UI:R), so it is not passively exploitable; impact is bounded to information disclosure from the queried page, manipulation of agent search results, and disruption of automation workflows rather than code execution. Publicly available exploit code exists, but there is no confirmed active exploitation (not in CISA KEV), and EPSS is low at 0.28% (19th percentile), making this a moderate, situational risk rather than a high-priority emergency.
The huggingface/transformers library, versions prior to 4.53.0, is vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically within the `normalize_numbers()` method of the `EnglishNormalizer`. Rated medium severity (CVSS 5.3), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically affecting the MarianTokenizer's `remove_language_code()` method. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability exists in the Hugging Face Transformers library, specifically in the `convert_tf_weight_name_to_pt_weight_name()` function. Rated medium severity (CVSS 5.3), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically within the DonutProcessor class's `token2json()` method. This vulnerability affects versions 4.50.3 and earlier, and is fixed in version 4.52.1. The issue arises from the regex pattern `<s_(.*?)>` which can be exploited to cause excessive CPU consumption through crafted input strings due to catastrophic backtracking. This vulnerability can lead to service disruption, resource exhaustion, and potential API service vulnerabilities, impacting document processing tasks using the Donut model.
Hugging Face Transformers versions up to 4.49.0 are affected by an improper input validation vulnerability in the `image_utils.py` file. The vulnerability arises from insecure URL validation using the `startswith()` method, which can be bypassed through URL username injection. This allows attackers to craft URLs that appear to be from YouTube but resolve to malicious domains, potentially leading to phishing attacks, malware distribution, or data exfiltration. The issue is fixed in version 4.52.1.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically in the `get_imports()` function within `dynamic_module_utils.py`. This vulnerability affects versions 4.49.0 and is fixed in version 4.51.0. The issue arises from a regular expression pattern `\s*try\s*:.*?except.*?:` used to filter out try/except blocks from Python code, which can be exploited to cause excessive CPU consumption through crafted input strings due to catastrophic backtracking. This vulnerability can lead to remote code loading disruption, resource exhaustion in model serving, supply chain attack vectors, and development pipeline disruption.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically in the `get_configuration_file()` function within the `transformers.configuration_utils` module. The affected version is 4.49.0, and the issue is resolved in version 4.51.0. The vulnerability arises from the use of a regular expression pattern `config\.(.*)\.json` that can be exploited to cause excessive CPU consumption through crafted input strings, leading to catastrophic backtracking. This can result in model serving disruption, resource exhaustion, and increased latency in applications using the library.
A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the huggingface/transformers repository, specifically in version 4.49.0. The vulnerability is due to inefficient regular expression complexity in the `SETTING_RE` variable within the `transformers/commands/chat.py` file. The regex contains repetition groups and non-optimized quantifiers, leading to exponential backtracking when processing 'almost matching' payloads. This can degrade application performance and potentially result in a denial-of-service (DoS) when handling specially crafted input strings. The issue is fixed in version 4.51.0.
A vulnerability in the `preprocess_string()` function of the `transformers.testing_utils` module in huggingface/transformers version v4.48.3 allows for a Regular Expression Denial of Service (ReDoS). Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file `tokenization_gpt_neox_japanese.py` of the. Rated medium severity (CVSS 6.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file tokenization_nougat_fast.py. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity.
Hugging Face Transformers Trax Model Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
Hugging Face Transformers MaskFormer Model Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
Hugging Face Transformers MobileViTV2 Deserialization of Untrusted Data Remote Code Execution Vulnerability. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, no authentication required, low attack complexity. No vendor patch available.
The huggingface/transformers library is vulnerable to arbitrary code execution through deserialization of untrusted data within the `load_repo_checkpoint()` function of the `TFPreTrainedModel()`. Rated critical severity (CVSS 9.6), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available.
Deserialization of Untrusted Data in GitHub repository huggingface/transformers prior to 4.36. Rated high severity (CVSS 7.8), this vulnerability is no authentication required, low attack complexity. Public exploit code available.
Deserialization of Untrusted Data in GitHub repository huggingface/transformers prior to 4.36. Rated high severity (CVSS 8.8), this vulnerability is remotely exploitable, low attack complexity. Public exploit code available.
Insecure Temporary File in GitHub repository huggingface/transformers prior to 4.30.0. Rated medium severity (CVSS 4.7). Public exploit code available.