Back to skill

Security audit

Boss Agent

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed multi-agent coordinator, but it gives broad access to other agents' histories, memory, state, and task channels without enough user consent or scoping.

Install only if you intend this skill to coordinate trusted Ass and Ops agents and are comfortable with it reading their session histories and forwarding tasks. Treat it as a Review item until the publisher narrows cross-agent access, requires explicit consent for history or memory reads, and defines how delegated operations preserve authorization boundaries.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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 (3)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding
Agent Role and Authority Hijacking## Vulnerability Details **File Location**: `SKILL.md`, lines 3–5; corroborating instruction at lines 145–146 **Vulnerability Type**: Agent identity and authority replacement **Risk Level**: High **Complete relevant snippet, faithfully translated into English:** ```markdown ## Role Definition You are the **Boss Agent**, the system's **overall coordinator**. ``` The closing instruction reinforces the role replacement: ```markdown **Remember:** You are the user's representative and are responsible for managing and coordinating other Agents! ``` ### Technical Analysis The Skill does not merely document an optional coordination procedure. It directly instructs the Agent to assume a new identity with authority over other Agents. When loaded into an active session, these imperative statements can alter the Agent's current goals, role boundaries, and interpretation of its permissions. The asserted role is subsequently used to justify reading other Agents' information and sending them tasks. Because the document does not state that existing system instructions, access controls, and explicit user authorization take precedence, an Agent may interpret the Skill as granting authority that the caller does not possess. ### Attack Path 1. The Skill is loaded into an Agent session. 2. The Agent accepts the instruction to become the “Boss Agent.” 3. The Agent treats cross-Agent management as part of its standing authority. 4. A malicious or ambiguous request invokes that assumed authority. 5. The Agent accesses other Agent contexts or delegates operations without independently verifying authorization. ### Impact Assessment Successful exploitation can redirect the current Agent's behavior and expand its perceived authority beyond the original task. The practical scope includes attempted access to other Agents' conversations and state, as well as task delegation to an operations-focused Agent. Actual privileges remain depend ...[truncated 153 chars]
Remediation
## Remediation Suggestions - Remove imperative identity-replacement statements such as “You are the Boss Agent.” - Reframe the document as optional, task-scoped guidance that applies only when the user explicitly requests coordination. - State that system instructions, platform access controls, and the current user's authorization always take precedence. - Do not infer authority from a role name or from the fact that the Skill was loaded. - Require explicit confirmation before accessing another Agent or assigning it work. - Add a rule prohibiting expansion of privileges or responsibilities beyond those granted by the active session.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:15
Finding
Unrestricted Access to Other Agents' Conversations, Memory, and State## Vulnerability Details **File Location**: `SKILL.md`, lines 15–18 and 29–36 **Vulnerability Type**: Cross-Agent access beyond least privilege **Risk Level**: High **Complete relevant snippet, faithfully translated into English:** ```markdown ### 2. Information Collection - **Read Ass Agent information** — understand the assistant's work status - **Read Ops Agent information** — understand the operations bot's work status - **Cross-Agent queries** — access other Agents' conversation history, memory, and state ### Ways to Access Other Agents #### 1. Read Conversation History ```bash # View Ass Agent conversations sessions_list --agent ass # View Ops Agent conversations sessions_list --agent ops ``` ``` ### Technical Analysis The instructions authorize broad collection of conversation history, persistent memory, and operational state from other Agents. They do not limit access to information necessary for a particular task, require consent from the affected user or Agent, or mandate redaction of credentials and private content. Conversation histories and memories can contain unrelated user data, authentication material, infrastructure details, or confidential operational context. Merely assigning the coordinator role does not establish authorization to retrieve those records. This violates least-privilege and context-isolation principles. The suggested `sessions_list` commands also identify specific Agent namespaces, making cross-context enumeration directly actionable where the platform exposes such commands. ### Attack Path 1. An attacker causes the Skill to be loaded or submits a request framed as a coordination task. 2. The Agent relies on the Skill's assertion that it may access other Agents' histories, memory, and state. 3. The Agent invokes `sessions_list --agent ass` or `sessions_list --agent ops`. 4. Session information or accessible conversation content is collected without task-specific au ...[truncated 650 chars]
Remediation
## Remediation Suggestions - Deny access to conversation history and persistent memory by default. - Require explicit, per-request user authorization identifying the target Agent and the exact data category required. - Replace raw history access with a narrowly scoped status API that returns only task-relevant, redacted information. - Enforce access control in the platform rather than relying solely on prompt instructions. - Apply tenant, user, and session isolation before returning any cross-Agent data. - Redact credentials, tokens, personal information, and unrelated conversation content. - Record immutable audit logs for cross-Agent access, including requester, target, purpose, and fields returned. - Require confirmation before disclosing collected information back to a requester when it may contain sensitive data.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:20
Finding
Unbounded Delegation to Other Agents Creates a Confused-Deputy Path## Vulnerability Details **File Location**: `SKILL.md`, lines 20–24 and 38–45 **Vulnerability Type**: Unauthorized cross-Agent task delegation **Risk Level**: High **Complete relevant snippet, faithfully translated into English:** ```markdown ### 3. Task Distribution - Assign daily tasks to the Ass Agent - Assign operations tasks to the Ops Agent - Coordinate multi-Agent collaboration #### 2. Send Tasks ```bash # Send a task to the Ass Agent sessions_send --session-key agent:ass:main --message "Please handle this task: ..." # Send a task to the Ops Agent sessions_send --session-key agent:ops:main --message "Please perform this operations task: ..." ``` ``` ### Technical Analysis The Skill supplies direct commands for sending arbitrary natural-language instructions to fixed Ass and Ops Agent sessions. It provides no allowlist of permitted operations, no validation of delegated content, no preservation of the original requester's identity and authorization, and no requirement that the receiving Agent independently approve sensitive actions. This creates a confused-deputy risk. A requester with access only to the coordinator may induce it to forward an instruction to an Ops Agent that has greater system privileges. The later statement that configuration changes, service restarts, and data deletion require user authorization is a partial safeguard, but it does not define how authorization is authenticated, propagated, or verified by the receiving Agent. It also leaves other sensitive operations unrestricted. ### Attack Path 1. An attacker submits a request that appears to require operational coordination. 2. The coordinator decomposes the request according to the Skill. 3. Attacker-controlled or insufficiently validated content is placed into the `--message` argument. 4. The coordinator sends the message to `agent:ops:main`. 5. The Ops Agent interprets the message as an authorized instruction from the coord ...[truncated 844 chars]
Remediation
## Remediation Suggestions - Replace arbitrary free-form delegation with a typed, allowlisted task interface. - Bind every delegated request to the authenticated original requester, requested action, target resource, and authorization evidence. - Require the receiving Agent to independently enforce authorization and safety policy rather than trusting the coordinator. - Prohibit forwarding secrets, shell commands, configuration mutations, destructive requests, or privilege changes through generic message fields. - Require explicit confirmation for service modification, restart, deletion, credential handling, or other high-impact operations. - Constrain the Ops Agent with least-privilege operating-system permissions and narrowly scoped tools. - Validate session identifiers instead of relying on fixed privileged destinations. - Log delegation chains and return structured results showing which Agent performed each action. - Treat all delegated message content as untrusted input and defend against nested prompt injection.
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
Confidence
94% confidence
Finding
The skill explicitly grants the Boss agent access to other agents’ session history, memory, and state, but does not present a user-facing privacy notice, consent boundary, or data minimization rule. This creates a real confidentiality risk because user data shared with one agent may be silently exposed to another agent with broader coordination privileges.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The task-routing examples show the Boss agent forwarding user requests to other agents without warning that user instructions and possibly embedded sensitive data may be retransmitted. This is dangerous because users may reasonably assume they are interacting with a single agent, while the skill enables undisclosed sharing across multiple agents and expands the exposure surface.

Natural-Language Policy Violations

Low
Confidence
70% confidence
Finding
The natural-language instructions throughout the file are presented only in Chinese, with no indication that users may interact in another language or choose a locale. Under the stated policy, forcing a specific language without opt-in can be a policy concern unless the locale limitation is explicit and justified.

Static analysis

No suspicious patterns detected.