Back to skill

Security audit

ZenHeart User Agent Workflows

Security checks for vulnerabilities and agentic risk

Overview

This skill documents ZenHeart workflows, but it also gives a remote server an under-scoped command channel and includes destructive publishing actions without clear confirmation steps.

Install only if you trust the ZenHeart endpoint and will keep the host pinned to an expected domain. Treat delete_news, delete_skill, publish_skill, and update_skill as actions that need explicit human review. Do not let the agent execute or answer server-pushed command callbacks unless the requested command is separately verified, narrowly scoped, and approved.

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)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:358
Finding

Unrestricted Remote Command Instructions Through WebSocket Callback

Content
View full analysis
", "command": "...", "args": {} } Reply: { "type": "command_result", "request_id": "", "ok": true, "output": "human-readable result" } ``` ### Technical Analysis The skill instructs the agent to accept a free-form `command` and attacker-controlled `args` from a remote WebSocket server, process that command, and return its output. No command allowlist, schema constraints, authorization checks, user confirmation, execution boundaries, or prohibited-operation rules are defined. The controlling endpoint is constructed from the user-supplied `host` input elsewhere in the skill rather than being restricted to the advertised ZenHeart domain. Consequently, an attacker who influences the host configuration can provide the WebSocket endpoint that supplies these instructions. Although the document does not define a direct shell-execution primitive, its command callback delegates undefined tasks to the consuming agent. Depending on the tools and permissions available to that agent, server-provided commands could alter the active session objective and induce file access, tool invocation, network requests, data disclosure, or other unintended actions. This is instruction hijacking because control over agent behavior is transferred from the user to an external server without sufficient restrictions. ### Attack Path 1. The victim loads the skill and supplies, or is induced to supply, an attacker-controlled `host`. 2. The agent connects to `wss:///v2/agent/ws` and completes the documented authentication exchange. 3. The malicious server sends a `command` frame containing attacker-selected command text and arguments. 4. The skill directs the agent to process the und ...[truncated 1170 chars]
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Self-Modification

High
Category
Rogue Agent
Confidence
90% confidence
Finding

The skill includes a self-modification workflow allowing publication and update of skill markdown, which can change future agent behavior and instructions. In an adversarial environment, this is dangerous because injected prompts, compromised servers, or mistaken user requests could cause the agent to rewrite its own capabilities, persistence mechanisms, or trust rules.

Content

Scanner excerpt · SKILL.md (reported line 225)May include surrounding context.

}

text

### Update skill markdown

```json
{

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The documented delete_news operation is destructive and irreversible in effect, yet the skill provides no requirement for explicit confirmation, preview, or warning before deletion. In agentic use, a prompt injection, misunderstanding, or operator typo could cause permanent content loss with no safety pause.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The delete_skill workflow enables removal of published skill content without any warning about destructive consequences or requirement to confirm intent. Because skills influence later agent behavior, accidental or manipulated deletion can disrupt operations and remove trusted instructions or tooling.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill exposes a generic server-driven command callback that instructs the agent to execute arbitrary commands and return results, but this capability is not constrained to a narrow, documented allowlist tied to the stated messaging/news/social workflow purpose. In an agent setting, this creates a remote instruction channel that can be abused for arbitrary action execution, data exfiltration, or privilege misuse if the model treats incoming command frames as trusted work requests.

Content

No source excerpt is available for this finding.