Back to plugin

Security audit

Tidegap — carbon-aware deferral

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked tidegap suspicious because it found a high-confidence authorization weakness that could expose or redirect users’ task results when the plugin is shared across users.

Consider installing only if the gateway and task queue are limited to one trusted user, or if the host independently enforces task ownership. In a shared deployment, the reported weakness could expose task results or redirect their delivery. The finding was confirmed in reviewed code, but exploitation depends on shared access without external ownership checks; isolated deployments are not shown to be affected. The verdict is suspicious, not a claim that the author intended harm.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

MCP07

Error
Location
plugin/dist/index.js:16410
Finding

Missing task-level authorization permits cross-user disclosure and delivery redirection on shared gateways ## Remediation Suggestions 1. **Persist authenticated ownership.** Associate every task and recurring job with a gateway-derived user or tenant identifier and its originating conversation. Do not accept ownership identities from model-controlled tool arguments. 2. **Enforce authorization on every operation.** Apply ownership checks to listing, detail retrieval, delivery changes, cancellation, deadline changes, replies, retries, and immediate execution. Filter records before returning them rather than exposing the shared queue and checking afterward. 3. **Authorize delivery changes atomically.** Require the target task to exist and belong to the caller before updating delivery settings. Perform the check and update transactionally where practical. Restrict administrative cross-user actions to explicit privileges. 4. **Bind chat delivery to the creator’s conversation.** Replace global `allowFrom[0]` selection with an authorized destination recorded at task creation. If no verified destination exists, retain the result in the queue instead of selecting another user. 5. **Handle legacy records conservatively.** Treat tasks without reliable ownership metadata as inaccessible to ordinary callers until an administrator explicitly assigns ownership or migrates them into an isolated queue. 6. **Constrain external delivery.** Consider administrator-approved webhook destinations and require authenticated approval for sensitive destination changes. Model-facing instructions to ask for confirmation are not a substitute for enforced authorization. 7. **Add multi-user regression tests.** Verify that one user cannot enumerate, read, redirect, cancel, or execute another user’s tasks. Test multiple Telegram recipients, shared CLI-created tasks, and legacy unowned records. 8. **Use deployment isolation as an interim mitigation.** Until ownership enforcement is implemente ...[truncated 129 chars]

Content
View full analysis

Vulnerability Details

File Location:

  • plugin/dist/index.js:16410–16429 — delivery configuration changes without task ownership verification.
  • plugin/dist/index.js:16225–16294 — shared queue enumeration and task-detail retrieval without caller-specific filtering.
  • plugin/dist/index.js:15229–15340 — global Telegram target selection and delivery configuration storage.

Vulnerability Type: Missing resource-level authorization; cross-session information disclosure.

Risk Level: High in multi-user deployments where mutually untrusted users share the plugin and queue without host-enforced isolation.

Vulnerable Code

The complete set_delivery execution handler validates delivery parameters but does not verify task existence or the caller’s authority over that task:

js
async execute(params, _config, context) {
  captureOpenClawRuntime(context);
  const cfg = {
    modes: params.deliver,
    webhookUrl: params.webhook_url,
    filePath: params.file_path,
    format: params.file_format
  };
  const error = validateDeliveryConfig(cfg);
  if (error) throw new Error(`set_delivery: ${error}`);
  await setDeliveryConfig(params.task_id, cfg);
  return { task_id: params.task_id, delivery: cfg };
}

Delivery configuration is stored solely under the supplied task ID:

js
async function setDeliveryConfig(taskId, config) {
  const existing = await readRecord(taskId);
  await writeRecord(taskId, { ...existing, ...config });
}

The scheduler’s persisted-task listing has no caller-specific filter:

js
listPersistedTasks() {
  return this.store?.list() ?? [];
}

Chat delivery uses the first entry in the gateway’s Telegram allowlist rather than a recipient associated with the task’s creator:

js
function telegramTarget(openclawConfig) {
  const tg = telegramConfig(openclawConfig);
  if (!tg?.enabled || !tg.botToken || !tg.allowFrom?.length) return void 0;
  return { token: tg.botToken, chatId: String(tg.allowFrom[0]) };
}

Techn

...[truncated 3643 chars]

SkillSpector was not run because this plugin release contains no bundled skills.

Endor dependency reachability

Checked Nov 12, 2026 · 8:36 PM UTC

No reachable vulnerabilities reported

Endor dependency scan completed.