Back to skill

Security audit

AgentSquared Skills

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent for AgentSquared messaging, but it asks agents to install or update a global runtime and restart a local gateway with broad authority.

Review this before installing if you are comfortable allowing AgentSquared to install/update a global npm CLI, run a persistent local gateway, use AgentSquared key files/tokens, and exchange or retrieve private friend messages. Prefer approving installs and updates explicitly, and use an isolated account or environment if possible.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:72
Finding

Unpinned Global Installation of a Security-Sensitive Runtime Dependency

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (9)

Self-Modification

High
Category
Rogue Agent
Confidence
90% confidence
Finding

The README instructs updating the local skills checkout via git pull and refreshing the CLI via npm, which are self-modification and software-change operations. In an agent setting, presenting these steps without strict confirmation and trust verification can lead an agent to alter its own operational environment, potentially introducing unreviewed code or breaking a working installation.

Content

Scanner excerpt · README.md (reported line 324)May include surrounding context.

Updating has two independent layers, but they should be checked together in normal operations:

Update Skills

bash
cd "<host-skills-root>/<checkout>"

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · README.md (reported line 416)May include surrounding context.

md
Typical examples:

- add a new workflow in `Skills` and also add new CLI support for it
- add a new host runtime in CLI and update skill docs to explain how to use it
- change the stable command surface in CLI and update human/agent docs in `Skills`

The rule is simple:

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 58)May include surrounding context.

md
- the current AgentSquared skill version from this root `SKILL.md`

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · SKILL.md (reported line 185)May include surrounding context.

md
Always translate CLI output for a non-technical owner. The owner only needs to understand that they are using the AgentSquared network, who they can talk to, what happened, and what they can ask next.

Default display rules:

- Keep it short, friendly, and action-oriented.
- Prefer `ready`, `needs setup`, `sent`, `received`, `unread`, `no friends yet`, or `ask me to send a message` over protocol details.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The README presents broad natural-language trigger examples like 'Check my AgentSquared setup' and 'Send a hello message' without strong activation boundaries or confirmation requirements. In an agentic environment, this can cause over-broad routing, leading the host agent to invoke networked or state-changing AgentSquared actions from casual user phrasing that may not reflect informed intent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This section documents commands that can send messages, show inbox contents, reveal conversation transcripts, restart gateways, and perform updates, but it does not pair those capabilities with clear privacy, consent, and system-impact warnings. If a host agent treats README guidance as operational instruction, a user may unknowingly trigger disclosure of private conversations or cause local system changes without informed approval.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill explicitly claims broad trigger phrases like 'update a2' or similar, which can cause it to activate on ambiguous owner requests and perform package/runtime update actions. In this context, activation leads to command execution that changes installed software and restarts a gateway, so over-broad routing increases the chance of unintended state-changing operations.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill’s selection guidance is broad and overlaps with common conversational intents such as greetings, check-ins, simple questions, and lightweight requests. In an agent environment, this can cause the skill to be selected in situations where a more constrained or purpose-built workflow should apply, increasing the chance of misrouting messages, bypassing intended review points, or handling sensitive interactions through an overly generic one-turn channel.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The sample conversation reports format time using the explicit locale/timezone "Asia/Shanghai" in both sender and receiver examples. Because this is natural-language presentation content in a general-purpose skill README, showing a fixed locale without opt-in can violate language/locale policy expectations.

Content

No source excerpt is available for this finding.