Back to skill

Security audit

JoinBot: your agent in Zoom, Meet and Teams, out loud

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners report concerns about unsolicited meeting contributions and saving meeting content to persistent memory; confidence is medium.

Before installing, consider whether unsolicited meeting speech and persistent storage of transcript-derived content fit your needs. The scan was partial: it inspected the listed Markdown component but could not resolve a local path-like reference, so the evidence is incomplete.

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
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (4)

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 52)May include surrounding context.

md
to address you. Calling it again updates your details. You can join once `status` is `approved`; if it is `pending`, tell your owner to approve you in the JoinBot app.
   - With a link: `join_meeting(link, bot_id)`.
   - For calendar meetings (events tagged `#joinbot`): `list_meetings()`, then take the `session_id` of the meeting whose status is `joined`.
2. Only if you have no wake-up webhook: call `listen(session_id)` in a loop **until it returns `{type: "ended"}`**. Don't stop on `idle`, and don't ask the user whether to continue; just call `listen` again.
3. On `{type: "question"}`: answer with `say(session_id, turn_id, spoken, chat?)` **within 8 seconds**. Answer fast: under 3 seconds it's spoken directly. At 3 seconds JoinBot says "Hmm, one sec." for you; past 8 seconds it tells the room the answer is coming in the chat, and your `say` is posted there instead.
4. To leave early when asked: `leave_meeting(session_id)`.
5. Speaking up uninvited: meetings default to `engagement: "he

Default engagement includes unsolicited speech rather than responding only when addressed by name.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The skill advertises answering when addressed by name, but defaults to unsolicited contributions based on meeting discussion and the agent's own knowledge. In a shared meeting, this can cause unexpected disclosure of private contextual information to attendees; frequency limits do not establish authorization to share it.

Content

No source excerpt is available for this finding.

Post-meeting instructions expand attendance into full-transcript retrieval and persistent memory retention.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The workflow automatically expands meeting attendance into retrieval of the full transcript and storage of selected content in persistent memory, without a separate authorization step. This collects and retains potentially sensitive statements from all participants beyond what is necessary to answer questions directed at the bot.

Content

No source excerpt is available for this finding.

The workflow saves meeting content to persistent memory without warning users that it can outlive the transcript's 24-hour deletion window.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

Saving transcript-derived content in agent memory creates a separate retained copy that is not covered by the stated 24-hour transcript deletion. Without disclosure or a retention policy for that memory, sensitive meeting information may remain available for later retrieval or unintended reuse; this overlaps with the preceding finding but specifically concerns retention transparency and lifecycle controls.

Content

No source excerpt is available for this finding.