llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · pydantic-ai

GHSA-5PPX-9JQ6-6PMW

Domain allowlists in web fetching and PlaywrightBrowser resolved refused hostnames before rejecting them

Texto original en inglés · Ver en español

Parche disponible
Descargar PNG JSON
CVSS 3.1
3.7
Baja
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:N/A:N Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
Parche2 versionesPrimer parche el mismo día de la divulgación
Exposición17 dDesde la divulgación, 4 nov 2026, 6:51 a. m.
Detección15 dDesde la divulgación hasta su ingreso al índice
Decisión estimadaSeguir
Remediar en la ventana regular de mantenimiento
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, no automatizable e impacto técnico parcial. Hay versiones corregidas disponibles.
Lo resuelve el mantenimiento de rutina.Esta no le quita el sueño ni a un SOC novato.
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada
Explotación
Remediación
Parche 2.66.0 +1
Monitoreo
Detectada +1
4 nov 26
20 nov 26
21 nov 26
4 nov 26
8 nov 26
13 nov 26
17 nov 26
21 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-5PPX-9JQ6-6PMW divulgada · CVSS 3.7 (GitHub)
Remediación Parche 2.66.0 pydantic-ai · 0 d desde divulgación
Remediación Parche 0.66.0 pydantic-ai-harness · 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

The local web-fetch tool and the `PlaywrightBrowser` harness capability looked up a hostname in DNS before checking it against their domain lists. A host outside the allowlist was refused and never contacted over HTTP, but its name still went to the resolver, which forwards it to that domain's nameserver. Data encoded in the hostname, such as `<data>.attacker.example`, could therefore leave the application despite the allowlist.

Details

The assumed attacker is a third party with no access to the application, who controls content the agent processes, such as a fetched page, a document, or a tool result. If that content induces the model to request a URL whose hostname encodes data from the conversation, the lookup delivers that data to the nameserver of a domain the attacker controls. Each lookup carries at most one hostname's worth of data, and exploitation depends on the injected instructions steering the model.

The lookup also made some refusals depend on DNS: in the web-fetch tool, a host outside the allowlist that resolved to a private address was reported as a private-address error rather than a domain-list refusal, revealing how the name resolved.

Domain lists are now checked before the hostname is resolved, for the initial request and every redirect in the web-fetch tool, and for every request the built-in `PlaywrightBrowser` `EgressPolicy` evaluates.

Who is affected

You are affected if the model can process content that untrusted parties influence, and either:

  • you configure `allowed_domains` on the local web-fetch tool, directly through `web_fetch_tool(allowed_domains=[...])` or through the `WebFetch` capability's local fallback (`pydantic-ai` / `pydantic-ai-slim`); or
  • you use `PlaywrightBrowser` (`pydantic-ai-harness`) with an `EgressPolicy` whose `allowed_domains` bounds every request kind, so that you rely on no request leaving for a host outside the list.

With `PlaywrightBrowser`'s default `allowlist_reach`, passive subresources such as images are not bounded by the allowlist, so a page can already signal to any host, as its documentation states. Configurations that use only `blocked_domains` were not relying on this boundary, because a blocklist permits every host it does not list.

Mitigation

Upgrade to a patched version of each affected package. If you use `PlaywrightBrowser` with an `EgressPolicy` subclass that overrides `refuse`, the patched version still resolves every host, because an override may permit hosts the domain lists refuse. Also override `needs_resolution` to return `False` for the hosts your override refuses by name. Until then, restrict DNS resolution for the process to the domains in your allowlist, for example with a resolver or egress policy that refuses other names.

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:H
BajaAlta
PrivilegiosPR:N
NingunoBajosAltos
InteracciónUI:N
NingunaRequerida
AlcanceS:U
CambiadoSin cambio
ConfidencialidadC:L
AltoBajoNinguno
IntegridadI:N
AltoBajoNinguno
DisponibilidadA:N
AltoBajoNinguno
Score de explotaciónQué tan fácil y probable es que la exploten
7.1/ 10Alto
Vector 7.1 (fácil) · sin evidencia de explotación.
Versiones

Rangos afectados

Afectada Corregida Última
pydantic-ai-harness≥ 0.52.0 ≤ 0.65.0
pydantic-ai · pydantic-ai-slim≥ 1.77.0 ≤ 1.107.9
pydantic-ai · pydantic-ai-slim≥ 2.0.0b1 ≤ 2.65.0
0.52.0
0.65.0
0.71.0
1.77.0
1.107.9
2.0.0b1
2.65.0
2.68.0
2.71.0
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
pypipydantic-ai
Corregida 2.66.0 0 d 2.68.0 no afectada
pypipydantic-ai-slim
Corregida 2.66.0 0 d 2.71.0 no afectada
pypipydantic-ai-harness
Corregida 0.66.0 0 d 0.71.0 no afectada
Versiones corregidas0.66.02.66.0
Explotación

Arsenal público

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

Referencias
Advisoryhttps://github.com/pydantic/pydantic-ai/security/advisories/GHSA-5ppx-9jq6-6pmw

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