Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-108957 Overview
CVE-2026-108957 is a code injection vulnerability [CWE-74] in TinyLLaVA_Factory, an open-source framework for small-scale large multimodal models. The flaw resides in the _load_text_config function within tinyllava/model/configuration_tinyllava.py, part of the TinyLlavaConfig component. Attacker-controlled input to the llm_model_name_or_path argument reaches a dynamic code path, resulting in code injection. The project uses a rolling release model, so no fixed version is available at publication time. A public exploit has been disclosed, and the maintainers have not yet responded to the issue report.
Critical Impact
Remote attackers can inject code through the llm_model_name_or_path configuration argument, though exploitation requires user interaction with the malicious configuration.
Affected Products
- TinyLLaVA TinyLLaVA_Factory up to commit 1abc6ac5e76a68d0fd087a071b56da968e69ca9b
- tinyllava/model/configuration_tinyllava.py module
- TinyLlavaConfig component (_load_text_config function)
Discovery Timeline
- 2026-10-13 - CVE-2026-108957 published to NVD
- 2026-10-15 - Last updated in NVD database
Technical Details for CVE-2026-108957
Vulnerability Analysis
The vulnerability affects the configuration loading logic of TinyLLaVA_Factory. The _load_text_config function processes the llm_model_name_or_path argument supplied to TinyLlavaConfig without adequate validation or sanitization. Because the value is used in a context that evaluates or imports attacker-influenced data, a crafted input triggers unintended code execution within the Python process loading the model. The attack vector is categorized as network-reachable, but exploitation requires user interaction — typically a user loading a malicious model configuration or repository reference.
The impact is bounded: confidentiality, integrity, and availability are each rated low in the CVSS 4.0 vector. However, because the injected code runs with the privileges of the user loading the model, downstream risk depends on the environment. Research environments with access to datasets, cloud credentials, or GPU infrastructure may face broader exposure.
Root Cause
The root cause is improper neutralization of special elements in data passed to a downstream component [CWE-74]. The _load_text_config routine trusts the llm_model_name_or_path parameter and resolves it in a way that permits code paths outside of pure string handling. No allow-list, path canonicalization, or type validation gates the value before it is consumed.
Attack Vector
An attacker distributes a crafted TinyLLaVA configuration, model card, or script that sets llm_model_name_or_path to a malicious value. When a victim loads the configuration using TinyLlavaConfig, the _load_text_config function processes the field and the injected payload executes. Delivery vectors include poisoned model repositories, pull requests, shared notebooks, and public datasets referencing the parameter. Refer to the GitHub Issue #212 Discussion and the VulDB CVE-2026-108957 Detail for additional context.
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-108957
Indicators of Compromise
- TinyLLaVA configuration files containing non-standard characters, shell metacharacters, or Python expressions in the llm_model_name_or_path field.
- Unexpected child processes spawned by Python interpreters that loaded TinyLLaVA model configurations.
- Outbound network connections from machine learning workstations immediately after model configuration loading.
Detection Strategies
- Scan source code repositories and model artifact stores for TinyLLaVA configuration files and inspect the llm_model_name_or_path values for anomalous content.
- Monitor Python processes importing tinyllava.model.configuration_tinyllava for unexpected subprocess execution, file writes, or network calls.
- Audit pip install and git clone activity targeting the TinyLLaVA_Factory repository to inventory exposed systems.
Monitoring Recommendations
- Enable process lineage telemetry on data science hosts to correlate model-loading activity with downstream execution.
- Alert on Python processes writing to startup paths, SSH authorized keys, or scheduled task locations following model configuration operations.
- Track access to cloud credential stores and GPU scheduler APIs by user accounts running TinyLLaVA workloads.
How to Mitigate CVE-2026-108957
Immediate Actions Required
- Treat all TinyLLaVA_Factory checkouts at or before commit 1abc6ac5e76a68d0fd087a071b56da968e69ca9b as vulnerable and restrict their use to isolated environments.
- Only load TinyLLaVA model configurations from trusted, internally reviewed sources until a patched release is available.
- Review existing configurations for suspicious llm_model_name_or_path values and remove any that reference untrusted paths or inline expressions.
Patch Information
The TinyLLaVA_Factory project follows a rolling release model, and no fixed version has been published. The maintainers were notified through an issue report but have not responded at the time of CVE publication. Monitor the GitHub Repository for TinyLLaVA and the VulDB Vulnerability #416776 entry for remediation updates.
Workarounds
- Run TinyLLaVA workloads inside containerized or sandboxed environments without access to production credentials or networks.
- Implement a wrapper around TinyLlavaConfig that validates llm_model_name_or_path against an allow-list of approved model identifiers before invocation.
- Apply egress filtering on hosts performing model experimentation to limit the impact of code execution originating from configuration loading.
# Example allow-list validation wrapper
ALLOWED_MODELS="TinyLlama/TinyLlama-1.1B TinyLLaVA/TinyLLaVA-Phi-2-SigLIP-3.1B"
if ! echo "$ALLOWED_MODELS" | grep -qw "$LLM_MODEL_NAME_OR_PATH"; then
echo "Rejected: $LLM_MODEL_NAME_OR_PATH is not an approved model identifier"
exit 1
fi
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

