EUVD-2026-17204
GHSA-gr6f-xx69-hrgq
Severity by source
AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
Primary rating from NVD · only source for this CVE.
CVSS VectorNVD
Lifecycle Timeline
3DescriptionCVE.org
In its design for automatic terminal command execution, HAI Build Code Generator offers two options: Execute safe commands and Execute all commands. The description for the former states that commands determined by the model to be safe will be automatically executed, whereas if the model judges a command to be potentially destructive, it still requires user approval. However, this design is highly susceptible to prompt injection attacks. An attacker can employ a generic template to wrap any malicious command and mislead the model into misclassifying it as a 'safe' command, thereby bypassing the user approval requirement and resulting in arbitrary command execution.
AnalysisAI
HAI Build Code Generator's automatic command execution feature can be bypassed through prompt injection attacks, allowing unauthenticated remote code execution by misleading the AI model into misclassifying malicious commands as safe. The vulnerability exploits a fundamental design flaw in the model's safety classification logic, where attackers can wrap destructive commands in generic templates to bypass the user approval requirement that should be triggered for potentially dangerous operations.
Technical ContextAI
HAI Build Code Generator implements an AI-driven command execution system that relies on language model inference to classify commands as either safe (auto-execute) or potentially destructive (require user approval). The vulnerability stems from the inherent susceptibility of large language models to prompt injection attacks-a class of attacks where specially crafted input prompts manipulate model behavior beyond intended parameters. The system's trust in model-based classification without cryptographic validation or sandbox isolation creates a single point of failure. This is a classic example of CWE-94 (Code Injection) and CWE-20 (Improper Input Validation), where the system fails to validate that user-supplied command templates are benign before submitting them to the AI model for classification.
RemediationAI
The primary remediation requires architectural redesign of the command execution safety mechanism. Instead of relying solely on AI model classification, implement a curated allowlist of genuinely safe commands that can be auto-executed, with all other commands requiring explicit user approval regardless of model judgment. For immediate risk reduction: disable automatic command execution entirely and require explicit user confirmation for all commands, or restrict command execution to a read-only sandbox environment where command output is captured but system-level side effects are prevented. Users should review the HAI Build project repository (https://github.com/presidio-oss/hai-build) and the LLM Tool-Calling CVEs tracking repository (https://github.com/Secsys-FDU/LLM-Tool-Calling-CVEs/issues/10) for any published patches or mitigation guidance. No vendor-released patch has been independently confirmed at the time of this analysis.
LibreNMS before 24.10.0 allows a remote attacker to execute arbitrary code via OS command injection involving AboutContr
Reachable assertion in Open5GS up to version 2.7.6 allows remote unauthenticated attackers to crash the Diameter CER han
A Command Injection issue in the payload build page in BYOB (Build Your Own Botnet) 2.0 allows attackers to execute arbi
Denial of service in Open5GS 2.7.6 and earlier allows network-adjacent attackers with low-privilege Diameter Gx peer acc
Denial of service in Open5GS HSS (Home Subscriber Server) versions 2.7.0 through 2.7.6 is triggered remotely by manipula
Remote command injection in Control Web Panel allows unauthenticated attackers to execute arbitrary OS commands as root
LINQPad before 5.52.01 Pro edition is vulnerable to Unsafe Deserialization in LINQPad.AutoRefManager::PopulateFromCache(
Remote code execution in dash-uploader (Python package for Plotly Dash) versions 0.1.0 through 0.7.0a2 allows unauthenti
Command injection in LiteLLM 1.18.10 lets attackers who can configure MCP (Model Context Protocol) servers supply arbitr
SQL injection in AzuraCast's backup restore flow (versions up to and including 0.23.7) lets an attacker achieve privileg
Arbitrary file read in Yu AI Code Mother v4.3 lets anonymous, remote attackers retrieve files outside the intended previ
Remote code execution in APScheduler (all versions through 3.10.x and 4.0.0a5) is achievable when applications deseriali
Same weakness CWE-94 – Code Injection
View allShare
External POC / Exploit Code
Leaving vuln.today