llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · memtomem-stm

GHSA-37R7-MCMJ-H6C2

Quadratic compression steps (HYBRID TOC fitting, repeated-heading keys, tail anomaly) let a crafted response stall the proxy event loop

Texto original en inglés · Ver en español

Sin parche publicado
Descargar PNG JSON
CVSS
—
high
Fuente: —
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
ParcheSin parcheEl fabricante no ha publicado corrección
Exposición29 dDesde la divulgación, 22 oct 2026, 5:36 p. m.
Detección29 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 la severidad declarada por GHSA (alta) y la evidencia de explotación (sin evaluación oficial): sin explotación conocida, no automatizable e impacto técnico parcial. No hay versión corregida: aplique mitigaciones.
Las PoC crecen rápido, como las llamas bebé.No es drama, es prevención. Hazme caso.
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada
Explotación
Remediación
Monitoreo
Detectada +1
22 oct 26
21 nov 26
22 oct 26
30 oct 26
6 nov 26
14 nov 26
21 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-37R7-MCMJ-H6C2 divulgada
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés

Summary

Three steps in memtomem-stm's compression tiers do quadratic or worse work in the size of an upstream tool response. These are algorithmic loops, not regexes. Compression runs synchronously on the proxy's event loop, so while one response is being compressed, every other call through the proxy waits.

The worst of them is reached under the default AUTO strategy by markdown with many headings: an 86 KB response stalls the proxy for about 37 s, and 118 KB of one repeated heading for about 73 s. `max_upstream_chars` defaults to 10,000,000, so it does not bound the input in practice.

Severity: High (availability). No data is exposed. The stall ends when compression finishes. No CVSS vector or CVE is asserted yet.

This is separate from GHSA-7jqx-w58w-8h4p, which covers quadratic regex scans on the same text path. Its fix does not touch these loops.

Affected steps (measured on main `5897eb75`, Python 3.13, `max_chars=8000`)

| step | what it does | crafted input | time |
|---|---|---|---|
| HYBRID TOC fitting (`HybridCompressor`, the TOC tail) | to fit the budget, drops TOC entries one at a time and binary-searches each preview, re-serializing the whole TOC with `json.dumps` at every step | markdown with distinct headings: 2,000 / 4,000 / 8,000 headings (20 / 42 / 86 KB) | 1.4 s / 7.9 s / 36.7 s |
| same, via repeated headings | same | 4,000 / 8,000 copies of one heading with distinct bodies (58 / 118 KB) | 16.3 s / 73.1 s |
| SELECTIVE `_unique_key` | for each repeated heading, probes the `#2`, `#3`, … suffixes from 2 upward | 4,000 / 8,000 / 16,000 copies of one heading (58 / 118 / 244 KB), parse alone | 0.5 s / 2.2 s / 9.8 s |
| TRUNCATE tail-anomaly path | builds the anomalous tail with `list.insert(0, …)`, which shifts every line already collected | `"prefix\n" * 2n + "tail\n" * n`, 1.2 / 2.4 / 4.8 MB | 0.44 s / 1.44 s / 4.8 s |

AUTO selects HYBRID for markdown with headings above the budget, and TRUNCATE for the repetitive plain text, so every row is reachable without configuration. SELECTIVE `_unique_key` is reached both through HYBRID and when SELECTIVE is configured directly.

Prerequisites and reach

  • The default configuration is affected.
  • The content must come from an upstream MCP server the operator configured. The realistic attacker is whoever controls content that such an upstream relays (a fetched web page, an issue body, a repository file), not the upstream's operator.

Fix

All three steps now run in linear time and return byte-identical output.

  • HYBRID measures candidate TOCs instead of serializing them. Each entry and the envelope are serialized once, the preview search keeps its probe sequence, and dropped entries are counted with prefix sums.
  • SELECTIVE keeps the next suffix per heading.
  • TRUNCATE collects the tail with `append` and reverses it once.

The fix is checked against the old code, used as an oracle, on random and shaped TOCs at every budget, and against seven mutations. On the inputs above the step now takes at most 0.08 s (it took 36.7 s / 73.1 s / 8.7 s before), and the tail path grows 4.0x for 4x input (11.5x before).

Discovery

Found during the review of the GHSA-7jqx-w58w-8h4p fix: the HYBRID and `_unique_key` steps by the plan review, the tail-anomaly path by the implementation review. All three were reproduced and measured independently.

Versiones

Rangos afectados

Afectada Corregida
memtomem-stm≤ 0.6.5
0.6.5
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
PyPImemtomem-stm
Sin parche — — —
Explotación

Arsenal público

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

Referencias
Advisoryhttps://github.com/memtomem/memtomem-stm/security/advisories/GHSA-37r7-mcmj-h6c2

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