CVE-2026-80047 Vulnerability Analysis & Exploit Details

CVE-2026-80047 Vulnerability Summary

CVE-2026-80047: A vulnerability in Hugging Face Transformers (versions >= 4.49.0 and <= 5.8.1) allows remote Python files to be written to local disk without user consent when using GenerativePreTrainedModel.load_custom_generate(). The function fetches and caches a remote module file before performing the required trust_remote_code consent check, inverting the security model enforced by other code-loading paths (such as AutoConfig, AutoModel, and AutoTokenizer). As a result, attacker‑controlled Python code from custom_generate/generate.py is copied into the user’s ~/.cache/huggingface/modules directory even if the user declines the trust prompt. Although execution is correctly gated, the file write is not reversible and can persist across sessions. This can lead to persistent, unauthorized files on disk and stale cache collisions where cached attacker code may later be executed during trusted model loads. The issue stems from an unconditional file write in dynamic_module_utils.py prior to any trust verification.

CVE-2026-80047 References

External References

• https://github.com/huggingface/transformers
• https://kb.cert.org/vuls/id/456290
• https://www.kb.cert.org/vuls/id/456290

CWE Common Weakness Enumeration

CAPEC Common Attack Pattern Enumeration and Classification

• Software Integrity Attack CAPEC-184 An attacker initiates a series of events designed to cause a user, program, server, or device to perform actions which undermine the integrity of software code, device data structures, or device firmware, achieving the modification of the target's integrity to achieve an insecure state.
• Malicious Software Download CAPEC-185 An attacker uses deceptive methods to cause a user or an automated process to download and install dangerous code that originates from an attacker controlled source. There are several variations to this strategy of attack.
• Malicious Software Update CAPEC-186 An adversary uses deceptive methods to cause a user or an automated process to download and install dangerous code believed to be a valid update that originates from an adversary controlled source.
• Malicious Automated Software Update via Redirection CAPEC-187 An attacker exploits two layers of weaknesses in server or client software for automated update mechanisms to undermine the integrity of the target code-base. The first weakness involves a failure to properly authenticate a server as a source of update or patch content. This type of weakness typically results from authentication mechanisms which can be defeated, allowing a hostile server to satisfy the criteria that establish a trust relationship. The second weakness is a systemic failure to validate the identity and integrity of code downloaded from a remote location, hence the inability to distinguish malicious code from a legitimate update.
• Malicious Manual Software Update CAPEC-533 An attacker introduces malicious code to the victim's system by altering the payload of a software update, allowing for additional compromise or site disruption at the victim location. These manual, or user-assisted attacks, vary from requiring the user to download and run an executable, to as streamlined as tricking the user to click a URL. Attacks which aim at penetrating a specific network infrastructure often rely upon secondary attack methods to achieve the desired impact. Spamming, for example, is a common method employed as an secondary attack vector. Thus the attacker has in their arsenal a choice of initial attack vectors ranging from traditional SMTP/POP/IMAP spamming and its varieties, to web-application mechanisms which commonly implement both chat and rich HTML messaging within the user interface.
• Open-Source Library Manipulation CAPEC-538 Adversaries implant malicious code in open source software (OSS) libraries to have it widely distributed, as OSS is commonly downloaded by developers and other users to incorporate into software development projects. The adversary can have a particular system in mind to target, or the implantation can be the first stage of follow-on attacks on many systems.
• Malicious Automated Software Update via Spoofing CAPEC-657 An attackers uses identify or content spoofing to trick a client into performing an automated software update from a malicious source. A malicious automated software update that leverages spoofing can include content or identity spoofing as well as protocol spoofing. Content or identity spoofing attacks can trigger updates in software by embedding scripted mechanisms within a malicious web page, which masquerades as a legitimate update source. Scripting mechanisms communicate with software components and trigger updates from locations specified by the attackers' server. The result is the client believing there is a legitimate software update available but instead downloading a malicious update from the attacker.
• Adversary in the Browser (AiTB) CAPEC-662 An adversary exploits security vulnerabilities or inherent functionalities of a web browser, in order to manipulate traffic between two endpoints.
• Spoof Open-Source Software Metadata CAPEC-691 An adversary spoofs open-source software metadata in an attempt to masquerade malicious software as popular, maintained, and trusted.
• Spoof Version Control System Commit Metadata CAPEC-692 An adversary spoofs metadata pertaining to a Version Control System (VCS) (e.g., Git) repository's commits to deceive users into believing that the maliciously provided software is frequently maintained and originates from a trusted source.
• StarJacking CAPEC-693 An adversary spoofs software popularity metadata to deceive users into believing that a maliciously provided package is widely used and originates from a trusted source.
• Repo Jacking CAPEC-695 An adversary takes advantage of the redirect property of directly linked Version Control System (VCS) repositories to trick users into incorporating malicious code into their applications.

Other 5 Recently Published CVEs Vulnerabilities

• CVE-2026-90555 – vLLM versions before 0.28.0 fail to validate audio sample rate headers in the transcription endpoint, allowing authenticated clients to bypass dura...
• CVE-2026-90554 – vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL m...
• CVE-2026-90553 – vLLM before 0.28.0 contains a remote code execution vulnerability in the LlavaOnevision2 processor loader that ignores the trust_remote_code parame...
• CVE-2026-90552 – WWBN AVideo through commit c3edcc274c389816d434acadac07ee78eaf330c1 fails to validate playlist ownership in the Playlists_schedules/list.json.php a...
• CVE-2026-90551 – WWBN AVideo through commit c3edcc274c389816d434acadac07ee78eaf330c1 fails to validate playlist ownership in the video_from_program API endpoint, al...