Back to skill

Security audit

Business Heartbeat Monitor

Security checks for vulnerabilities and agentic risk

Overview

This skill is a business monitoring template, but it gives agents broad automatic authority over production recovery, process termination, customer replies, and payment retries without enough scoping or approval gates.

Review this skill before installing in any real business environment. Use it only with explicit allowlists, least-privilege read-only access by default, human approval for production changes, customer messages, and payment retries, and clear logging, retry limits, and rollback procedures.

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

Error
Location
SKILL.md:19
Finding

Unrestricted Autonomous Operational Actions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:19-64
Vulnerability Type: T05: Unauthorized Access and Privilege Escalation
Risk Level: High

Vulnerable Instructions

markdown
2. If any fail:
   - Attempt restart (via deployment redeploy, service restart, etc.)
   - If auto-recovery works: log to daily notes, move on
   - If auto-recovery fails: alert immediately on primary channel
3. Log all results to daily notes
markdown
2. If a service is down:
   - Attempt restart
   - If restart fails: alert with error details
3. If a process has been stalled for 2+ consecutive heartbeats:
   - Kill and restart automatically
   - Log the stall duration and any error output
markdown
3. For routine questions (order status, how-to, account access):
   - Draft a response
   - Send if within Tier 1 autonomy, otherwise flag for review
4. For complex issues:
   - Summarize the problem
   - Flag for human review with recommended action
markdown
1. Check for failed payments or declined charges
2. Flag any for retry or customer follow-up
3. Track daily revenue running total
4. Note any anomalies (unusual spikes, drops, large transactions)
markdown
1. Check for any listed active processes (tmux sessions, workers, agents)
2. For each: verify it's still running
3. If dead: restart automatically, log what happened
4. If stalled (same output for 2+ checks): kill and restart
5. If completed successfully: report completion, clean up

Technical Analysis

The Skill instructs an agent to perform consequential state-changing operations automatically, including deployment redeployment, service restart, process termination, customer communication, and potentially payment retries. These instructions do not define or enforce:

  • An allowlist of resources that the agent may modify.
  • Least-privilege credentials for each monitored system.
  • Human approval for destructive or financially significant actions.
  • Reliable criteria for disti ...[truncated 2637 chars]
Remediation
View remediation

Remediation Suggestions

  1. Default to observation-only behavior

    • Make health checks, inbox scans, and payment checks read-only by default.
    • Require an explicit policy grant before enabling any state-changing operation.
  2. Enforce resource allowlists

    • Enumerate the exact services, processes, deployment targets, inboxes, and payment accounts the agent may access.
    • Reject operations against resources not present in the configured allowlist.
  3. Require approval for high-impact actions

    • Require human confirmation before redeploying production services, terminating processes, sending customer messages, or retrying payments.
    • Permit automatic restart only for specifically approved, stateless resources with tested recovery procedures.
  4. Apply least-privilege credentials

    • Use separate credentials for monitoring, service recovery, support communication, and payment operations.
    • Do not grant write permissions to an agent whose configured task requires only status checks.
    • Restrict credentials by environment, resource, operation, and time where supported.
  5. Harden failure detection

    • Use multiple independent signals before declaring a resource dead or stalled.
    • Distinguish unchanged output from actual process failure.
    • Add minimum observation windows, retry limits, cooldown periods, and circuit breakers.
  6. Protect financial operations

    • Require human approval for payment retries.
    • Use provider-supported idempotency keys and verify transaction state before retrying.
    • Define maximum retry counts and monetary thresholds.
  7. Protect outbound communications

    • Treat all inbox content as untrusted data rather than instructions.
    • Restrict automated responses to approved templates and narrowly defined scenarios.
    • Prevent messages from causing tool invocation or changes to monitoring policy.
  8. Add recovery and accountability controls

    • Record the triggering evidence, aut ...[truncated 447 chars]
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 (4)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill authorizes the agent to kill and restart processes based on a simplistic 'stalled for 2+ checks' heuristic, with no requirement for human review, safe-stop logic, state preservation, or workload awareness. This is dangerous because false positives could terminate live jobs, corrupt in-flight work, or repeatedly flap critical background systems, causing availability and integrity issues.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill explicitly instructs the agent to attempt service restarts and redeploy-style recovery during routine monitoring, but it does not require confirmation, guardrails, rollback criteria, or user-visible warnings about the operational risk. Automated recovery can worsen incidents, interrupt healthy-but-degraded services, or create cascading failures if health checks are wrong or the restart action is mis-scoped.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs the agent to scan inbox contents and draft or send customer responses without privacy notices, data-handling constraints, or strict autonomy boundaries. This creates risk of unauthorized access to sensitive communications, accidental disclosure of personal data, and inappropriate outbound messaging if the agent misclassifies an issue as routine.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill directs the agent to inspect failed payments, declined charges, and revenue anomalies, but provides no warning about sensitive financial data handling, least-privilege access, or restrictions on taking action. In practice, this could expose payment-related data unnecessarily or lead the agent to trigger customer follow-up or retries without sufficient controls.

Content

No source excerpt is available for this finding.