Neo4jChatAgent passes LLM-generated Cypher queries straight to the Neo4j driver with no validation, no statement-type allowlist, and no opt-out gate. The query text is influenceable by prompt injection (direct user input or indirect content the agent reads back via RAG), so an attacker who can influence the prompt can read or destroy all graph data and, when APOC or dbms.security procedures are enabled on the server, achieve OS-command and filesystem access. This is the same defect class and threat model as the SQLChatAgent prompt-to-SQL-to-RCE issue fixed in version 0.63.0 (CVE-2026-25879); that fix did not extend to the neo4j module. ## Technical detail Untrusted-input to sink trace (reviewed on langroid HEAD b9df06f, v0.65.3): 1. Tool schemas accept raw query text from the LLM. `langroid/agent/special/neo4j/tools.py:4-9` (CypherRetrievalTool.cypher_query: str) and `:15-21` (CypherCreationTool.cypher_query: str). These tools are enabled unconditionally in `neo4j_chat_agent.py:412-419` (enable_message([GraphSchemaTool, CypherRetrievalTool, CypherCreationTool, DoneTool])). 2. Read path. `neo4j_chat_agent.py:300` cypher_retrieval_tool(msg) -> `:325` query = msg.cypher_query -> `:328` self.read_query(query) -> `:223` session.run(query, parameters). The LLM-controlled string is the first positional argument to session.run; parameters is None. No validation occurs between :325 and :223. 3. Write path. `neo4j_chat_agent.py:338` cypher_creation_tool(msg) -> `:348` query = msg.cypher_query -> `:351` self.write_query(query) -> `:276` session.write_transaction(lambda tx: tx.run(query, parameters)). The only inspection of the string (write_query :251-264) is a query.upper() substring test for CREATE/MERGE/CONSTRAINT/INDEX whose sole effect is setting self.config.database_created = True; it blocks nothing. A query such as `MATCH (n) DETACH DELETE n` (the same statement the built-in remove_database helper runs at :287-293) passes unrestricted. 4. Guarded-sibling contrast proving the incomplete fix. The SQLChatAgent, patched for the parent CVE, validates every query before execution. `langroid/agent/special/sql/sql_chat_agent.py:256` defines allow_dangerous_operations: bool = False (opt-in gate); `:618` _validate_query runs (a) a dangerous-pattern regex blocklist (_DANGEROUS_SQL_PATTERNS, :628-641), (b) a sqlglot statement-type allowlist defaulting to SELECT-only (:643-686), and (c) an AST-level dangerous-function blocklist (:687-704). run_query calls rejection = self._validate_query(query) at `:721` before executing. The neo4j read_query/write_query paths have no equivalent: a grep for _validate_query/allow_dangerous in neo4j_chat_agent.py returns nothing. The defense exists for SQL and is simply absent for Cypher, which is the definition of an incomplete fix for the same prompt-to-query-language injection class. ## Proof of concept (static, no third-party systems, no live Neo4j required) Step 1 (presence vs absence of the guard, one command): ``` grep -n "_validate_query\|allow_dangerous" langroid/agent/special/sql/sql_chat_agent.py # SQL: many hits (gate + validator + call site :721) grep -n "_validate_query\|allow_dangerous" langroid/agent/special/neo4j/neo4j_chat_agent.py # neo4j: ZERO hits ``` Step 2 (sink trace is direct, no interposed check): ``` read_query: msg.cypher_query (line 325) -> read_query(query) (328) -> session.run(query, parameters) (223) write_query: msg.cypher_query (348) -> write_query(query) (351) -> tx.run(query, parameters) (276) ``` The only string inspection (write_query 251-264) is a .upper() substring test that only sets database_created=True and rejects nothing. The asymmetry (validator + opt-out gate enforced on every SQL query at sql_chat_agent.py:721; nothing on either Cypher path) is the deterministic artifact: the same project, for the same injection class, guards one query language and not the other. A dynamic confirmation against an operator-owned disposable Neo4j (create a CSPRNG-marked node via cypher_creation_tool, read it back via cypher_retrieval_tool, then DETACH DELETE it) reproduces the read/write/destroy primitive without any third-party system. ## Impact An attacker who can influence the agent prompt (directly, or indirectly via content the agent reads back through RAG) controls the executed Cypher. Floor (no extra config, not contingent): unauthorized read of all graph data via cypher_retrieval_tool and full write/destroy (including MATCH (n) DETACH DELETE n) via cypher_creation_tool, plus the built-in LOAD CSV remote-fetch (SSRF) primitive. Ceiling (config-conditional, when APOC / dbms.security procedures are granted to the DB role, a common deployment): apoc.load.* (SSRF + remote/local file read), apoc.cypher.runFile / apoc.import.* / apoc.export.* (filesystem), and CALL dbms.* admin procedures, i.e. the Cypher analogue of the parent's COPY ... FROM PROGRAM RCE primitive. This mirrors the privileged-role contingency the parent advisory (CVE-2026-25879) accep

AI Analysis