CVE-2026-54769
CriticalCVSS 10.0Exploitation Probability (EPSS)
Elevated risk56th percentile - higher than 56% of all known CVEs
Summary
Langroid versions prior to 0.65.2 contain a critical sandbox escape vulnerability leading to Remote Code Execution (RCE) in its `TableChatAgent` and `VectorStore` components. When evaluating LLM-generated tool messages with `full_eval=True`, the `eval()` function fails to remove `__builtins__` from the globals dictionary, allowing access to system functions like `__import__('os').system()`. An attacker can exploit this by providing a crafted prompt to achieve unauthenticated RCE on the host.
Risk Assessment
The organization is at risk of complete system compromise by an unauthenticated attacker, potentially leading to data theft, malware installation, or service disruption.
Recommendation
Immediately upgrade Langroid to version 0.65.2 or later. If upgrading is not possible, disable the `full_eval=True` feature in `TableChatAgent` and `VectorStore` components.
Other vulnerabilities in Langroid
See all- CVE-2026-55615Critical
Langroid before version 0.65.5 has a vulnerability in the Neo4jChatAgent component that allows Cypher query injection. An attacker can influence the query via prompt injection, enabling reading or destroying all graph data, and with APOC or dbms.security procedures enabled, achieving OS command and filesystem access.
- CVE-2026-54760Critical
In Langroid framework prior to version 0.65.1, a SQL injection vulnerability was found in the `SQLChatAgent` component. Despite default dangerous operations being disabled, the protection mechanism using a regex blocklist and SELECT allowlist can be bypassed with crafted PostgreSQL queries, enabling server file reads.
- CVE-2026-25879Critical
Langroid SQLChatAgent before version 0.63.0 executes SQL generated by an LLM, which is influenceable by prompt injection. An attacker can coerce execution of dangerous operations like COPY ... FROM PROGRAM, leading to remote code execution on the database host.
- CVE-2026-54771High
A vulnerability in Langroid framework prior to version 0.65.3 allows untrusted users to directly invoke tools via raw JSON payloads, even when tools are registered with `use=False, handle=True`.
- CVE-2026-50181High
In the Langroid framework, the ReadFileTool and WriteFileTool do not validate that the file path stays within the working directory (curr_dir). This allows an attacker to read and write files outside the allowed directory using path traversal sequences like ../.
- CVE-2026-50180High
In Langroid library prior to version 0.64.0, the SQLChatAgent component has an imperfect _validate_query defense layer that fails to block all dangerous SQL primitives. The blocklist misses PostgreSQL file-reading functions (pg_read_file, pg_stat_file) and SQL Server/SQLite constructs, allowing an attacker to read files from the database host.
Original NVD description (English source)
Langroid is a framework for building large-language-model-powered applications. Versions prior to 0.65.2 are vulnerable to a critical Sandbox Escape leading to Remote Code Execution (RCE) in its `TableChatAgent` and `VectorStore` capabilities. When these agents evaluate LLM-generated tool messages with `full_eval=True`, they attempt to sandbox the execution by explicitly setting `locals` to an empty dictionary `{}` inside Python's `eval()` function. However, this relies on an incomplete understanding of Python's execution model. Because `__builtins__` is not explicitly scrubbed from the `globals` dictionary mapping, Python implicitly injects all built-ins during execution, granting full access to functions like `__import__('os').system()`. Since `TableChatAgent.pandas_eval()` executes external LLM outputs natively, this bypass permits any attacker providing prompt payload to achieve unauthenticated RCE on the host system. Version 0.65.2 patches the issue.