Back to skill

Security audit

Meeting Autopilot

Security checks for vulnerabilities and agentic risk

Overview

This meeting assistant is coherent and not malicious, but it asks the agent to automatically read broad calendar, email, document, conversation, and meeting-memory data with weak scoping and retention controls.

Review this skill carefully before installing. It may be useful for work meetings, but only enable it in an agent environment where you can limit calendar, email, document, and memory access; exclude private or sensitive events; approve what gets stored; and delete retained meeting notes and action items.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
skill.md:33
Finding
Overbroad Automatic Data Access and Indefinite Retention Without Explicit Authorization Controls## Vulnerability Details **File Location**: `skill.md:33`, `skill.md:45-50`, `skill.md:104-105`, `skill.md:115`, `skill.md:173-176`, and `skill.md:219` **Vulnerability Type**: T05: Unauthorized Access and Privilege Escalation **Risk Level**: Medium ### Vulnerable Instructions `skill.md:33`: ```markdown Trigger: "Prep me for my meeting with X", or automatically 30 minutes before any calendar event ``` `skill.md:45-50`: ```markdown - Recent email threads or conversations with each attendee - Any open action items or commitments between you and them - Last time you spoke and what was discussed **Context and preparation** - Relevant documents or emails related to the meeting topic ``` `skill.md:104-105`: ```markdown - Stored in your agent's memory for future meeting prep - Searchable: "What did we decide about X in our last call with Y" ``` `skill.md:115`: ```markdown - Every action item from every meeting is stored with owner, task, deadline, and source meeting ``` `skill.md:173-176`: ```markdown - Every processed meeting is stored and fully searchable - Ask natural questions and the agent retrieves relevant context - Cross-references people, topics, decisions, and dates - Surfaces institutional memory that would otherwise be lost in old notes ``` `skill.md:219`: ```markdown No configuration required. The agent reads your calendar automatically. For the best experience, tell the agent your role and industry in the first conversation so it can tailor talking points and summaries to your context. ``` ### Technical Analysis The skill directs the agent to read calendar information automatically and retrieve related emails, documents, conversations, relationship history, commitments, and meeting records. It then instructs the agent to retain every processed meeting and action item in searchable memory. These instructions do not define explicit authorization checks for each con ...[truncated 2945 chars]
Remediation
## Remediation Suggestions 1. Require explicit user opt-in before connecting to calendar, email, document, conversation, or persistent-memory services. 2. Disable automatic pre-meeting processing by default and require the user to enable it for selected calendars or meetings. 3. Use minimum OAuth or API scopes, such as read-only calendar access, and request additional sources separately only when required. 4. Restrict retrieval to user-selected accounts, folders, participants, date ranges, and meeting identifiers. 5. Display the sources and records selected for retrieval before incorporating sensitive content into a brief. 6. Require separate approval before persisting meeting notes, action items, attendee details, or retrieved communications. 7. Establish configurable retention periods and provide clear view, export, correction, and deletion controls. 8. Encrypt persisted meeting data and apply authorization checks to every memory query. 9. Prevent retrieval across unrelated meetings, organizations, clients, or user profiles through tenant and context isolation. 10. Avoid attendance and missed-commitment profiling unless the user separately enables that feature and has a legitimate basis to process the information. 11. Record auditable logs showing which source was accessed, why it was accessed, and what information was retained. 12. Replace the “No configuration required” statement with an explicit setup process describing requested permissions, privacy effects, retention behavior, and revocation procedures.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (8)

Missing User Warnings

High
Confidence
96% confidence
Finding
The skill describes collecting calendar details, attendee backgrounds, relationship history, recent emails, prior conversations, and related documents, but it does not present a prominent upfront warning about this breadth of data access. In a meeting-assistant skill, silent or understated access to such communications and interpersonal context materially raises privacy, consent, and over-collection risks.

Missing User Warnings

High
Confidence
97% confidence
Finding
The skill states that meeting summaries are stored in memory and reused for search and future preparation, but it does not provide a clear warning about persistent retention of notes, decisions, attendee information, and commitments. Persistent storage of workplace discussions can create substantial confidentiality and compliance risk if users do not understand retention scope, discoverability, and future reuse.

Vague Triggers

Medium
Confidence
95% confidence
Finding
The skill auto-activates 30 minutes before any calendar event and pulls meeting context without clear scoping, exclusions, or explicit per-meeting consent. In a meeting-assistant context this can expose sensitive attendee, email, and prior-conversation data for personal, confidential, or irrelevant calendar events, and can trigger on events the user did not intend the agent to process.

Vague Triggers

Medium
Confidence
92% confidence
Finding
Using a common phrase like "Meeting starting now" as a trigger is ambiguous and can be invoked during ordinary conversation or forwarded text, causing unintended note-taking mode activation. In this skill, accidental activation is more dangerous because the agent may start capturing sensitive meeting content or commitments without the user's deliberate intent.

Vague Triggers

Medium
Confidence
90% confidence
Finding
Automatically offering follow-up behavior on a generic phrase like "Meeting done" creates an unnecessary risk of unintended workflow execution and downstream disclosure of meeting summaries. Although the skill says emails are not sent without approval, accidental triggering can still generate drafts or expose stored notes at the wrong time.

Vague Triggers

Medium
Confidence
94% confidence
Finding
Automatic action-item tracking after every meeting lacks boundaries on what meetings are in scope and what data may be persisted. In this context, broad automatic persistence can capture commitments, names, deadlines, and relationship details from sensitive meetings and retain them beyond what the user expected.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The setup section casually states that the agent reads the user's calendar automatically, without a prominent explanation of scope, permissions, or privacy implications. This is risky because automatic calendar access can reveal sensitive schedules, private events, locations, invitees, and meeting metadata even before any meeting-specific command is issued.

Vague Triggers

Low
Confidence
88% confidence
Finding
Auto-activating recurring-meeting intelligence based on a broad heuristic such as repeated attendees can infer long-term behavioral patterns without explicit consent. While useful functionally, it increases privacy risk by building attendance and performance profiles from routine meetings that may not be appropriate to analyze automatically.

Static analysis

No suspicious patterns detected.