llamaCVEVulnerability Intelligence
/
Advisory · GitHub · npm · knowns

GHSA-H73X-698R-QRVG

Multi-Stage Workspace Context Escalation & Sandbox Escape via Bootstrap Gate Bypass and Unbounded Directory Enumeration

Parche disponible
Descargar PNG JSON
CVSS 4.0
9.4
Crítica
CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
Parche1 versiónPrimer parche a 13 d de la divulgación
Exposición73 dDesde la divulgación, 8 sept 2026, 11:00 p. m.
Detección71 dDesde la divulgación hasta su ingreso al índice
Decisión estimadaSeguir de cerca
Monitorear y programar la remediación
Estimación de Llama con el árbol SSVC de CISA a partir de vector CVSS 4.0 y la evidencia de explotación (sin evaluación oficial): sin explotación conocida, no automatizable e impacto técnico total. Hay versiones corregidas disponibles.
Esto en un informe de auditoría se ve feísimo.Tu firewall merece enterarse de esto.
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada +1
Explotación
Remediación
Parche 0.34.0
Monitoreo
Detectada +1
9 sept 26
22 sept 26
19 nov 26
21 nov 26
9 sept 26
27 sept 26
15 oct 26
3 nov 26
21 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-H73X-698R-QRVG divulgada · CVSS 9.4 (GitHub)
Divulgación Modificada Última actualización del registro
Remediación Parche 0.34.0 knowns · 13 d desde divulgación
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés

Overview

Verified. A two-primitive exploit chain in Knowns allows a low-privilege AI agent operating under a sandboxed `PresetReadOnly` session to escape its designated project root, escalate to `CapAdmin`, and assume full read/write/code-mutation access over arbitrary target repositories on the host filesystem.

The chain links an unauthenticated directory enumeration endpoint (`GET /api/workspaces/browse`) that discloses valid Knowns project roots anywhere on the host, with a statically hardcoded bootstrap gate bypass in `NewGuardMiddleware` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L25-L64 (`internal/permissions/guard.go:25-64` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L25-L64) that unconditionally exempts `project.set` from capability enforcement regardless of session state. Because the bootstrap exemption carries no precondition check on whether a store is already initialized, an already-constrained agent calls `project.set` without touching the capability gate, swaps `s.store` and `s.root` to an arbitrary target, and immediately inherits that target's permission policy - which defaults to `PresetReadWriteNoDelete`, granting write, generate, archive, and code-mutation access.

The exploit requires no race condition, no privileged account, and no user interaction. It executes deterministically over HTTP and MCP from any network-reachable low-privilege session.

Affected Paths

| File Path | Role | Vulnerability & Execution Impact |
| :--- | :--- | :--- |
| `internal/server/routes/workspace.go:41-110` https://github.com/knowns-dev/knowns/tree/main/internal/server/routes/workspace.go#L41-L110 | Directory Enumeration Endpoint | `browse()` https://github.com/knowns-dev/knowns/tree/main/internal/server/routes/workspace.go#L41-L110 - Unbounded host filesystem traversal (CWE-22): `os.ReadDir(filepath.Abs(dir))` accepts arbitrary `?path=` values with no allowlist or root-containment check. Returns absolute host paths with `isProject: true` markers, disclosing the location of any valid Knowns project on the host. |
| `internal/permissions/guard.go:25-64` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L25-L64 | Permission Enforcement Middleware | `NewGuardMiddleware()` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L25-L64 - Bootstrap exemption without state precondition (CWE-863): `isBootstrapAction(toolName, action)` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 at line 34 returns `true` for `project.set` regardless of whether an active store is already initialized, bypassing `CheckCapability` unconditionally. |
| `internal/permissions/guard.go:107-113` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 | Bootstrap Action Classifier | `isBootstrapAction()` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 - Static allowlist, no state guard: Hardcodes `project.set`, `project.detect`, `project.current`, `project.status` as always-exempt. No check against `getStore() == nil`. |
| `internal/permissions/registry.go:53` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/registry.go#L53 | Capability Registry | `project.set` classified as `CapAdmin / RiskMedium`: All standard presets (including `PresetReadOnly`, `PresetReadWrite`, `PresetReadWriteNoDelete`, `PresetGenerateDryRun`) explicitly deny `CapAdmin`. The bootstrap exemption is the only path past this denial. |
| `internal/mcp/handlers/project.go:186-213` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/handlers/project.go#L186-L213 | Project Switch Handler | `handleProjectSet()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/handlers/project.go#L186-L213 - Accepts arbitrary `projectRoot`, no containment check: Validates `.knowns/config.json` existence only. Passes the caller-supplied path directly to `setStore()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269. |
| `internal/mcp/server.go:245-269` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269 | Active Context State | `setStore()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269 - Atomic store and root swap under mutex: Replaces `s.store`, `s.root`, and `s.lspManager` in a single lock acquisition. All subsequent tool calls operate against the new root and its policy. |
| `internal/mcp/server.go:322-332` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L322-L332 | Permission Config Loader | `permConfigLoader()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L322-L332 - Dynamic policy reload on every call: Reads `cfg.Settings.Permissions` from the current `s.store` on each MCP tool invocation. A store swap immediately changes the active policy for the entire session. |

Root Cause

The bootstrap exemption in `internal/permissions/guard.go:34-36` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L34-L36 was designed to allow `project.set` during the initial server startup window when no store exists and the permission system cannot yet load its config. This is architecturally valid for startup. The flaw is that `isBootstrapAction` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 carries no runtime state precondition - it returns `true` regardless of whether `s.store` is `nil` or fully initialized. A session already constrained by `PresetReadOnly` can therefore invoke `project.set` at any time without the call touching `CheckCapability`.

Flaw A - Bootstrap Exemption Without State Guard (`guard.go:34-36`)

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/permissions/guard.go#L32-L36

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/permissions/guard.go#L107

`project.set` is classified in `internal/permissions/registry.go:53` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/registry.go#L53 as `CapAdmin / TargetRuntime / RiskMedium`. Every standard permission preset denies `CapAdmin`:

  • `PresetReadOnly` (`policy.go:50-62`): `CapAdmin: true` in `Denied`
  • `PresetReadWrite` (`policy.go:63-76`): `CapAdmin: true` in `Denied`
  • `PresetReadWriteNoDelete` (`policy.go:77-89`): `CapAdmin: true` in `Denied`
  • `PresetGenerateDryRun` (`policy.go:90-104`): `CapAdmin: true` in `Denied`

Because `isBootstrapAction` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 short-circuits before `CheckCapability` is ever reached, the denial in every preset is never evaluated.

Flaw B - Unbounded Directory Enumeration (`workspace.go:41-110`)

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/server/routes/workspace.go#L41-L62

`internal/server/routes/workspace.go:80-86` https://github.com/knowns-dev/knowns/tree/main/internal/server/routes/workspace.go#L80-L86 (project detection logic):

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/server/routes/workspace.go#L80-L86

No allowlist, no root containment, no authentication check on the `?path=` parameter. The endpoint walks arbitrary host directories and returns absolute paths annotated with `isProject: true` wherever a `.knowns/config.json` is found. An agent with only HTTP access to port 4141 can enumerate the entire host filesystem for valid switch targets.

Flaw C - Dynamic Policy Swap via `setStore` + `permConfigLoader` (`server.go:245-332`)

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/mcp/server.go#L245-L250

https://github.com/knowns-dev/knowns/blob/c6ed4cb8548c266930910d3b24b7605fa2c21150/internal/mcp/server.go#L322-L332

`permConfigLoader` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L322-L332 is closed over `getStore()`, which always reads the live `s.store` under a read lock. The moment `setStore` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269 completes, every subsequent MCP call evaluates permissions against the target repository's config. If the target has no explicit `permissions` config (the common case), `EffectivePolicy` falls back to `DefaultPreset` = `PresetReadWriteNoDelete`, granting `CapRead`, `CapWrite`, `CapGenerate`, and `CapArchive` - and access to `code.replace` / `code.insert` / `code.rename` (`RiskHigh` write operations).

Attack Phases

| Phase | Step | Effect |
| :--- | :--- | :--- |
| 1. Enumerate | `GET /api/workspaces/browse?path=/home/victim/projects` | Response discloses `{"path": "/home/victim/projects/financial-services-backend", "isProject": true}` |
| 2. Escalate | MCP tool call: `project` / `action: set` / `projectRoot: /home/victim/projects/financial-services-backend` | `isBootstrapAction("project","set")` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 returns `true`. `CheckCapability` is never called. `handleProjectSet` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/handlers/project.go#L186-L213 validates `.knowns/` presence and calls `setStore()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269. |
| 3. Swap | `setStore()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269 atomically replaces `s.store`, `s.root`, `s.lspManager` | Session now operates in target repository under target's policy (defaults to `PresetReadWriteNoDelete`) |
| 4. Mutate | `code` / `action: replace` targeting `auth/tokens.py` | `ensureCodeMutationPath` checks containment against `s.root` = `/home/victim/projects/financial-services-backend`. Code edit proceeds. |

Proof of Concept

Phase 1 - Target Discovery
GET /api/workspaces/browse?path=/home/victim/projects HTTP/1.1
Host: 127.0.0.1:4141

Response:

[
  {
    "name": "financial-services-backend",
    "path": "/home/victim/projects/financial-services-backend",
    "isProject": true,
    "hasChildren": true
  }
]
Phase 2 - Context Escalation via Bootstrap Bypass
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "project",
    "arguments": {
      "action": "set",
      "projectRoot": "/home/victim/projects/financial-services-backend"
    }
  }
}

Execution trace:

  1. `NewGuardMiddleware` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L25-L64 intercepts the call.
  2. `isBootstrapAction("project", "set")` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go#L107-L113 evaluates to `true` at `guard.go:34`.
  3. `CheckCapability` is never reached. `CapAdmin` denial in the active preset is never evaluated.
  4. `handleProjectSet` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/handlers/project.go#L186-L213 validates `.knowns/config.json` in the target path.
  5. `setStore()` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L245-L269 acquires the write lock and sets `s.store` and `s.root` to the target.
Phase 3 - Post-Escalation Code Mutation
{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "code",
    "arguments": {
      "action": "replace",
      "path": "auth/tokens.py",
      "search": "def verify_token(token):\n    return jwt.decode(token, SECRET_KEY)",
      "replace": "def verify_token(token):\n    return True"
    }
  }
}

`ensureCodeMutationPath` verifies the path is contained within `s.root` = `/home/victim/projects/financial-services-backend`. The store swap satisfies the check. The code mutation proceeds.

Impact

  • Confidentiality: Critical - Full read access across all files, tasks, documents, and memory entries in any target repository the agent discovers via `browse`.
  • Integrity: Critical - Full write, code-mutation, and deletion access in target repositories after escalation. Includes `code.replace`, `code.insert`, `code.rename` - all `RiskHigh` write operations - and doc/task/memory mutation.
  • Availability: High - `code.delete` and `docs.hard_delete` (`CapDelete, RiskHigh`) become reachable once the target's policy is loaded, enabling destruction of source files, task records, and documentation.
  • Subsequent System Impact: High - The agent escapes the confines of its initial read-only sandbox and operates with the permissions of an arbitrary target project on the same host.
  • No user interaction required - The chain executes programmatically with no victim action beyond running the Knowns MCP server.

CVSS v4.0

CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H
Score: 9.4 Critical

Rationale for each metric:

| Metric Group | Metric | Value | Rationale |
| :--- | :--- | :--- | :--- |
| Exploitability | Attack Vector (AV) | Network (N) | Reached over HTTP (browse endpoint) and MCP over WebSocket/SSE. No local access required. |
| | Attack Complexity (AC) | Low (L) | Deterministic two-step sequence. No race condition, no timing dependency, no brute force. |
| | Attack Requirements (AT) | None (N) | Default API posture allows both browse and MCP session connection. |
| | Privileges Required (PR) | Low (L) | Requires a valid low-privilege MCP session (e.g. `PresetReadOnly` guest agent). |
| | User Interaction (UI) | None (N) | Executes programmatically without victim action after session establishment. |
| Vulnerable System | VC (Confidentiality) | High (H) | Full read access across files, tasks, documents, and memory in the current workspace. |
| | VI (Integrity) | High (H) | Full modification of workspace tasks, docs, and code upon bypass. |
| | VA (Availability) | High (H) | Full deletion and destruction capability in target workspace. |
| Subsequent System | SC (Confidentiality) | High (H) | Full read access across arbitrary separate repositories and projects on host. |
| | SI (Integrity) | High (H) | Arbitrary code mutation in external host repositories. |
| | SA (Availability) | High (H) | Code deletion and project destruction across external host repositories. |

Remediation

Fix 1 - Gate Bootstrap Exemption on Uninitialized State (`guard.go` https://github.com/knowns-dev/knowns/tree/main/internal/permissions/guard.go)

Bootstrap exemptions must only apply when no active store exists. Once a session has an initialized store, `project.set` is an administrative context-switch and must pass through `CheckCapability`:

--- a/internal/permissions/guard.go
+++ b/internal/permissions/guard.go
@@ -25,7 +25,7 @@ func NewGuardMiddleware(configLoader ConfigLoader) server.ToolHandlerMiddleware {
-func NewGuardMiddleware(configLoader ConfigLoader) server.ToolHandlerMiddleware {
+func NewGuardMiddleware(configLoader ConfigLoader, isStoreInitialized func() bool) server.ToolHandlerMiddleware {
     return func(next server.ToolHandlerFunc) server.ToolHandlerFunc {
         return func(ctx context.Context, req gomcp.CallToolRequest) (*gomcp.CallToolResult, error) {
             toolName := req.Params.Name
             args := req.GetArguments()
             action, _ := args["action"].(string)

             // Bootstrap actions are only allowed when no project context is yet established.
-            if isBootstrapAction(toolName, action) {
+            if isBootstrapAction(toolName, action) && !isStoreInitialized() {
                 return next(ctx, req)
             }

Pass `func() bool { return getStore() != nil }` as `isStoreInitialized` at the call site in `internal/mcp/server.go:340` https://github.com/knowns-dev/knowns/tree/main/internal/mcp/server.go#L340.

Fix 2 - Restrict `browse` to Allowlisted Workspace Roots (`workspace.go` https://github.com/knowns-dev/knowns/tree/main/internal/server/routes/workspace.go)

The `browse` endpoint must confine traversal to an explicit allowlist of authorized parent directories. Arbitrary `?path=` values must be rejected:

--- a/internal/server/routes/workspace.go
+++ b/internal/server/routes/workspace.go
@@ -52,6 +52,9 @@ func (wr *WorkspaceRoutes) browse(w http.ResponseWriter, r *http.Request) {
     absDir, err := filepath.Abs(dir)
     if err != nil {
         respondError(w, http.StatusBadRequest, "invalid path")
         return
     }
+    if !wr.isAllowedBrowsePath(absDir) {
+        respondError(w, http.StatusForbidden, "path outside allowed workspace roots")
+        return
+    }

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

Vector CVSS 4.0

Desglose de métricas

Opción resaltada = valor del vector · izquierda = peor caso · GitHub
Vector de ataqueAV:N
RedAdyacenteLocalFísica
ComplejidadAC:L
BajaAlta
Requisitos de ataqueAT:N
NingunoPresentes
PrivilegiosPR:L
NingunoBajosAltos
InteracciónUI:N
NingunaPasivaActiva
Confid. (sistema)VC:H
AltoBajoNinguno
Integridad (sistema)VI:H
AltoBajoNinguno
Disponib. (sistema)VA:H
AltoBajoNinguno
Confid. (posterior)SC:H
AltoBajoNinguno
Integridad (posterior)SI:H
AltoBajoNinguno
Disponib. (posterior)SA:H
AltoBajoNinguno
Score de explotaciónQué tan fácil y probable es que la exploten
8.3/ 10Alto
Vector 8.3 (muy fácil) · sin evidencia de explotación.
Versiones

Rangos afectados

Afectada Corregida Última
knowns≤ 0.33.0
0.33.0
0.35.0
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
npmknowns
Corregida 0.34.0 13 d 0.35.0 no afectada
Versiones corregidas0.34.0
Explotación

Arsenal público

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

Referencias
Advisoryhttps://github.com/knowns-dev/knowns/security/advisories/GHSA-h73x-698r-qrvg

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