CVE-2026-68770
CriticalCVSS 9.8Exploitation Probability (EPSS)
Low risk49th percentile - higher than 49% of all known CVEs
Summary
sentence-transformers contains a security control bypass vulnerability in the import_module_class helper in util/misc.py. The condition 'or os.path.exists(model_name_or_path)' allows code execution from Python files placed in a local model directory even when trust_remote_code=False. An attacker who can influence the model directory contents can place malicious files that execute when the model is loaded.
Risk Assessment
The organization is at risk of arbitrary code execution when loading models with SentenceTransformer if an attacker can control the model directory contents. This could lead to application compromise or data leakage.
Recommendation
Update the sentence-transformers library to a version that fixes this vulnerability. Until then, avoid loading models from untrusted locations and verify model directory contents.
Related vulnerabilities
- CVE-2026-86106Critical
An unauthenticated actor with network access to the private HA interconnect may trigger sensitive HA peer functions without verification. This could result in elevated command execution on Edge units where HA is enabled.
- CVE-2026-73453Critical
An unauthenticated P4Runtime client can achieve arbitrary code execution under certain conditions on affected platforms running Arista EOS configured with P4Runtime. By crafting a malicious packet during session initiation, an attacker can obtain complete administrative control over the switch.
- CVE-2026-81642Critical
In NLnet Labs Unbound up to and including 1.26.0, a vulnerability in the DNSSEC validator enables denial of service and possible remote code execution via DNSKEY processing. A crafted DNSKEY can overflow the digest buffer.
- CVE-2026-27565Critical
An unauthenticated remote attacker can upload a malicious IODD file that places and executes a shell script with root privileges. The script remains active even after a reboot.
- CVE-2026-27546Critical
An unauthenticated remote attacker can exploit an authentication bypass in the _account_log function to log in as an admin, even when accounts are properly configured.
- CVE-2026-73447Critical
A privileged attacker can exploit gNSI Certz service on Arista EOS-based products to escalate privileges and execute arbitrary OS commands via a crafted Certz Rotate request. The Bootz service is also affected.
- CVE-2026-14349Critical
The TrueBooker – Appointment Booking and Scheduler System plugin for WordPress in all versions up to and including 1.2.3 is vulnerable to authorization bypass. This is due to the plugin not properly verifying that a user is authorized to perform an action. This allows unauthenticated attackers to modify the email address of arbitrary user accounts, including administrators, which can be leveraged to reset the account password and gain access.
- CVE-2026-12793Critical
The JetFormBuilder — Dynamic Blocks Form Builder plugin for WordPress in all versions up to and including 3.6.2 is vulnerable to Privilege Escalation. This is due to the plugin not validating that a submitted form ID belongs to a JetFormBuilder form before parsing the referenced post's content as form schema and executing an Advanced Validation server-side callback. This allows unauthenticated attackers to create a new administrator-level user account.
- CVE-2026-15640Critical
Under certain conditions a valid SAML IdP response may be used to impersonate another Secret Server user. The vulnerability affects the SAML authentication mechanism.
- CVE-2026-15639Critical
An attacker can craft a malicious link that, if used by a legitimate user, may cause the user's browser to run JavaScript supplied by the attacker. The vulnerability affects the link processing mechanism.
Original NVD description (English source)
sentence-transformers contains a security control bypass vulnerability that allows attackers to achieve arbitrary code execution by exploiting a logic flaw in the import_module_class helper within sentence_transformers/util/misc.py, where the guard condition includes an 'or os.path.exists(model_name_or_path)' clause that satisfies the trust gate whenever the supplied path exists on the local filesystem, regardless of the trust_remote_code=False argument. Attackers who can control or influence the contents of a model directory on disk can place malicious Python files such as modeling_*.py referenced via modules.json, causing the code to execute at import time when an application loads the model with SentenceTransformer(path, trust_remote_code=False), bypassing the documented security contract and achieving code execution within the loading process.