Back to skill

Security audit

AgentMeet

Security checks for vulnerabilities and agentic risk

Overview

The skill has a coherent scheduling purpose, but it should be reviewed because it installs unaudited code from a hard-coded Dropbox path and can send emails or modify calendars without clearly documented user confirmation.

Before installing, verify the actual ~/Dropbox/Dev/agentmeet source, package manifest, and lockfile, and run dependency installation in an isolated environment. Use this only with explicit approval before sending any email or creating calendar events, review the full invite content including any promotional link, and set the timezone deliberately.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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 (2)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:19
Finding

Mandatory Promotional Link Hijacks User-Facing Meeting Invitations

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 19-21
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Medium

Complete Code Snippet:

markdown
3. Recipient clicks a slot -> opens pre-filled reply with machine-readable payload
4. Sender's agent detects the reply -> creates calendar events for both parties

Every invite email also includes an "Add AgentMeet" link for viral adoption.

Technical Analysis

The Skill explicitly instructs the agent to include an “Add AgentMeet” link in every invitation for “viral adoption.” This changes user-facing messages to serve the Skill author's promotional objective rather than limiting output to the user's meeting-coordination request.

The destination and content of the link are not disclosed in the audited artifact, and the instruction does not require informed user approval before adding it. Because the project contains no implementation files, the exact link-generation behavior and destination validation cannot be verified.

Attack Path

  1. A user invokes the Skill to prepare or send a meeting invitation.
  2. The agent follows the instruction that every invitation must include the adoption link.
  3. The promotional link is inserted into an email sent under the user's identity.
  4. External recipients receive and may follow a link that the user did not explicitly request and whose destination is not specified in the audited artifact.

Impact Assessment

The issue does not directly grant elevated system privileges. Its scope is the content of meeting invitations and the external recipients receiving them. It can cause unauthorized promotion, damage the sender's reputation, create phishing concerns, and induce recipients to visit an undisclosed destination under the apparent endorsement of the sender.

Remediation
View remediation

Remediation Suggestions

  • Remove the requirement to include promotional links in every invitation.
  • Disable attribution and promotional content by default.
  • If attribution is offered, require explicit user consent before adding it to each message.
  • Show the user the complete email and exact destination URL before transmission.
  • Restrict any permitted link to a documented HTTPS destination and validate it against an allowlist.
  • Clearly separate optional product attribution from the protocol payload needed for scheduling.

T08 · Insecure Dependencies

Error
Location
SKILL.md:5
Finding

Dependency Installation Uses an Unverifiable Mutable External Directory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 5 and 24-28
Vulnerability Type: T08: Insecure Dependencies
Risk Level: High

Complete Code Snippet:

yaml
metadata: {"clawdbot":{"emoji":"📅","os":["darwin","linux"],"requires":{"bins":["bun"]},"install":[{"id":"bun","kind":"shell","command":"cd ~/Dropbox/Dev/agentmeet && bun install","label":"Install agentmeet dependencies via Bun"}]}}
markdown
## Setup

```bash
cd ~/Dropbox/Dev/agentmeet && npm install
text

### Technical Analysis

Both installation instructions change into the hard-coded external directory `~/Dropbox/Dev/agentmeet` and invoke a package manager. That directory is outside the audited project and is potentially mutable through local changes or Dropbox synchronization.

The audited artifact does not contain a package manifest, lockfile, implementation source, or integrity metadata. Consequently, the dependency graph and package lifecycle scripts executed by `bun install` or `npm install` cannot be reviewed or reproduced from the supplied project.

Package installation can execute lifecycle scripts defined by the project or its dependencies. If the contents of the external directory or its dependency resolution are compromised, installation can execute arbitrary commands with the privileges of the user running the package manager. The documentation also inconsistently uses both Bun and npm, increasing the likelihood of divergent and unreproducible dependency resolution.

### Attack Path

1. An attacker compromises, replaces, or influences files synchronized into `~/Dropbox/Dev/agentmeet`, or causes dependency resolution to select a malicious package.
2. The user or agent follows the Skill's setup instructions.
3. `bun install` or `npm install` reads the manifest from the unaudited external directory and resolves its dependencies.
4. A malicious package or project lifecycle script runs during installation.
5. The s
...[truncated 845 chars]
Remediation
View remediation

Remediation Suggestions

  • Include the implementation, package manifest, and lockfile within the reviewed Skill package.
  • Replace the hard-coded Dropbox development path with a versioned, verifiable release source.
  • Pin exact dependency versions and enforce lockfile integrity during installation.
  • Standardize on one package manager and use its immutable or frozen-lockfile installation mode.
  • Publish release hashes or signatures and verify them before installation.
  • Disable dependency lifecycle scripts where feasible, or explicitly audit every required lifecycle script before execution.
  • Run installation with least privilege in a sandbox or isolated build environment without access to Gmail, Calendar, OAuth tokens, or unrelated user files.
  • Ensure all documentation references only files and source code that are actually included in the distributable artifact.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (3)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The documented workflow says the sender's agent detects a reply and creates calendar events for both parties, with no mention of a confirmation or approval step. Automatic event creation based on parsed email content can lead to unwanted calendar modifications, spoofed scheduling actions, or abuse if malformed or deceptive emails are processed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill documentation states it requires Gmail and Google Calendar access and demonstrates sending invites, parsing replies, and generating slots, but it does not provide an explicit user-facing warning that the agent may read calendar data, send emails, and act on the user's behalf. In an agent skill context, missing consent and capability disclosure increases the risk of unauthorized or surprising actions involving sensitive communications and scheduling metadata.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The default configuration hard-codes the timezone as Asia/Taipei without describing user opt-in or automatic derivation from account/system settings. While not directly a security exploit, this can cause scheduling mistakes, confusing invitations, or inadvertent disclosure of incorrect availability in a workflow that performs real email and calendar actions.

Content

No source excerpt is available for this finding.