llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · praisonaiagents

GHSA-9426-Q567-PC36

Security hardening across servers, file and network access in PraisonAI 4.8.0

Sin parche publicado
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
ParcheSin parcheEl fabricante no ha publicado corrección
Exposición1 dDesde la divulgación, 19 nov 2026, 12:34 p. m.
Detección4 minDesde la divulgación hasta su ingreso al índice
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada +1
Explotación
Remediación
Monitoreo
Detectada +1
12:34 p. m.
02:39 p. m.
12:34 p. m.
07:05 p. m.
01:36 a. m.
08:07 a. m.
02:39 p. m.
FechaFaseEventoDetalle
Divulgación Publicada GHSA-9426-Q567-PC36 divulgada · CVSS 9.8 (GitHub)
Divulgación Modificada Última actualización del registro
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés

Summary

PraisonAI 4.8.0 is a security hardening release. It tightens defaults and adds shared checks across eight areas: network servers, the gateway and bots, outbound fetching, file paths, project-file trust, approvals, sandboxes, and input validation. Earlier versions could, in some default or documented configurations, expose agent tools to other hosts, read or write files outside the working folder, fetch internal addresses, or honour settings from an untrusted project folder.

Users should upgrade every installed PraisonAI package to the patched versions below.

Affected packages

| Ecosystem | Package | Affected versions | Patched versions |
|---|---|---|---|
| pip | `praisonaiagents` | `< 1.8.0` | `1.8.0` |
| pip | `praisonai` | `< 4.8.0` | `4.8.0` |
| pip | `praisonai-code` | `< 0.2.0` | `0.2.0` |
| pip | `praisonai-bot` | `< 0.2.0` | `0.2.0` |
| pip | `praisonai-mcp` | `< 0.2.0` | `0.2.0` |
| pip | `praisonai-sandbox` | `< 0.2.0` | `0.2.0` |
| pip | `praisonai-browser` | `< 0.2.0` | `0.2.0` |
| pip | `praisonai-deploy` | `< 0.2.0` | `0.2.0` |
| npm | `praisonai` | `< 1.8.0` | `1.8.0` |

Details by area

Servers and binds. Some built-in servers listened on all network interfaces, or accepted requests without a token, in their default or documented configuration. Some keyless local servers also accepted cross-site browser requests, and a few compared tokens in non-constant time or kept unbounded request state. Every server now binds `127.0.0.1` by default, refuses a public bind without a token, checks Origin and Host when keyless, caps request bodies, and compares tokens in constant time.

Gateway and bots. Scoped gateway tokens could reach routes and sessions beyond their scope, and some bot interactions (buttons, slash commands, email senders, shared-channel approvals) were not limited to allowed users. Routes and sessions now follow token scopes, and bot interactions follow the same allowlist as messages.

Outbound fetching. Several URL-reading paths did not apply, or did not pin, the private-address check. They now share one fetch helper that refuses private, loopback and metadata addresses and pins the checked address. Credentials from the environment are sent only to their configured origin.

File paths. Several file, search, recipe and bot helpers did not confine paths to the working folder. They now share one containment check that resolves symlinks.

Project-file trust. A project folder's config file could start MCP servers, load tool code, or set capability-granting options such as the model endpoint or shell expansion. These now count only from the user's own config, the environment or code.

Approvals. Some tools that send messages, email or schedules ran without approval, and some grants were broader than intended. These tools now need approval by default, grants match their arguments, and MCP tool filters apply at call time.

Sandboxes and child environment. Some sandbox backends accepted policies they could not enforce, and model-run commands inherited the full process environment. Unenforceable policies are now refused, and child processes get only an allowlisted environment.

Input validation. Some values reached SQL, IMAP, deployment commands, generated files or browser scripts without validation. They are now validated or passed as literals.

No proof-of-concept steps are included. Reports with the same root cause were merged into this advisory.

Severity

Overall: Critical, 9.8, `CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H`.

This is our own score, validated against the default configuration of the affected versions. It is the highest area: a built-in server reachable on the network without a token, where impact depends on the tools the user registered. The other areas range lower:

| Area | Highest default-configuration score |
|---|---|
| Servers and binds | 9.8 Critical |
| Project-file trust | 8.8 High |
| Input validation | 7.7 High |
| File paths | 7.5 High |
| Outbound fetching | 7.4 High |
| Gateway and bots | 6.5 Medium |
| Approvals | 6.5 Medium |
| Sandboxes and child environment | opt-in configurations only (up to 7.5 High) |

Weaknesses: CWE-1188 (insecure default), CWE-306 (missing authentication), CWE-346 (origin validation), CWE-22 (path traversal), CWE-918 (server-side request forgery), CWE-829 (inclusion from an untrusted control sphere), CWE-862 (missing authorization), CWE-693 (protection mechanism failure), CWE-74 (injection), CWE-770 (allocation without limits), CWE-208 (timing discrepancy).

Patches

Upgrade to `4.8.0`:

pip install -U praisonai          # upgrades every Python package listed above
npm install praisonai@latest

Some defaults change. The release notes list every change and the setting that adjusts it: https://github.com/MervinPraison/PraisonAI/releases/tag/v4.8.0

Workarounds

If you cannot upgrade yet: bind servers to `127.0.0.1` or set a token on each one; do not run PraisonAI inside a folder you do not trust; do not set `ALLOW_LOCAL_CRAWL`; and keep approval on for tools that send messages or run code.

Credits

We thank everyone who reported these issues privately:

@1chig0, @1diot9, @3rdn4Li, @a25370, @alexps9, @Alien0ne, @anantchakravarthy, @arpitjain099, @avrlab233, @baeseungwon1010, @BarakSrour, @baseforce65, @btnalit, @c240030, @cazy8, @ChineseNopa, @dajiaohuang, @daridor9, @DennyDai, @devzephyr, @duccuongst, @hgrgr, @hsnyus-09, @kta1kri, @manus-pi, @manus-use, @niuzhouxin, @node9ai, @rajivraj, @rampage0010, @rexpository, @romain-deperne, @sec-reex, @Seven11Eleven, @SnailSploit, @sour-exploit, @superbot-helper, @superboy-zjc, @SuyueGuo, @tonghuaroot, @Vancir, @venkatesh2003631, @vinhluongptit, @vndasunkid, @vndasunkid11, @W-Rookie, @wleeaf, @wzunknown, @YashvantHange, @zachariahchow-gvt, @zhunlian, @ziyue-pan, @zohuyhieuzo03, @zt2misay2

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
Versiones

Rangos afectados

Afectada Corregida Última
praisonai-code · praisonai-bot · praisonai-mcp · praisonai-sandbox · praisonai-browser · praisonai-deploy< 0.2.0
praisonaiagents · praisonai< 1.8.0
praisonai< 4.8.0
0.2.0
1.7.22
1.8.0
4.7.26
4.8.0
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
pypipraisonaiagents
Sin parche — — 1.7.22 afectada
pypipraisonai
Sin parche — — 4.7.26 afectada
Versiones corregidas0.2.01.8.04.8.0
Explotación

Arsenal público

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

Referencias
Advisoryhttps://github.com/MervinPraison/PraisonAI/security/advisories/GHSA-9426-q567-pc36

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