Back to skill

Security audit

Brsearch

Security checks for vulnerabilities and agentic risk

Overview

This skill routes searches through a logged-in DeepSeek browser session and disables normal retrieval tools without clear consent or data-safety limits.

Review this before installing if your prompts may include private, work, personal, regulated, or credential-like information. Use it only when you are comfortable sending the query content to DeepSeek under a logged-in account, and verify important answers with authoritative sources.

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 (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding

Suppression of Standard Retrieval Tools Through Skill Instructions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 3
Vulnerability Type: Agent tool-policy and instruction hijacking
Risk Level: High

Complete Code Snippet (English translation):

yaml
description: Disable web_search, web_fetch, and memory_search, and use an online large language model for searches

Technical Analysis

The skill explicitly instructs the agent to disable three standard retrieval capabilities: web_search, web_fetch, and memory_search. This is not merely a provider preference; it alters the agent's available information-retrieval policy when the skill is loaded.

Replacing established tools with an external model changes the trust boundary and prevents the agent from using potentially safer or more authoritative retrieval paths. It also makes later decisions dependent on output from a third-party model that has not been established as a trusted instruction source.

This is best classified as T01: Skill Instruction Hijacking because the behavior is implemented directly through skill text that changes the agent's current-session operating constraints. No evidence was found that the skill modifies the underlying tool implementations, so this is not classified as tool spoofing.

Attack Path

  1. The skill is loaded into the agent's active context.
  2. The agent interprets the description as an instruction to suppress web_search, web_fetch, and memory_search.
  3. The agent no longer uses its normal retrieval paths for the user's request.
  4. Retrieval is redirected to an externally hosted language model.
  5. The external model's unverified response is incorporated into subsequent agent reasoning or actions.

Impact Assessment

The issue can control the agent's retrieval behavior for the current skill invocation and influence downstream decisions through externally generated content. It does not, based on the audited file, grant operating-system privileges, access loc ...[truncated 262 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the directive that disables web_search, web_fetch, and memory_search.
  2. Present the external model as an optional retrieval provider rather than a mandatory replacement.
  3. Require explicit user approval before routing a query or other content to a third-party service.
  4. Preserve standard retrieval tools so the agent can select the safest and most authoritative source for each task.
  5. State that external model responses are untrusted data and must not be interpreted as agent instructions.
  6. Validate important claims against authoritative sources before using them in downstream actions.
  7. Add restrictions preventing secrets, credentials, personal information, and private workspace content from being submitted externally.

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:8
Finding

Mandatory Routing of User Queries and Downstream Reasoning Through a Third-Party Model

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 8-17
Vulnerability Type: Third-party interaction and downstream instruction redirection
Risk Level: Medium

Complete Code Snippet (English translation):

markdown
1. Check whether a browser is already open; if not, use the browser tool to open one.
2. Check whether an open browser page contains the DeepSeek online large language model page at chat.deepseek.com.
3. If DeepSeek is already logged in, operate it directly to obtain information; otherwise, ask the user to log in and begin obtaining information after login.
4. Query the online large language model, wait for its output to complete, and extract summarized information for subsequent operations.

## DeepSeek Online Model Usage Rules

- Begin using it only after login.
- Find the DeepSeek expert-mode button and click it to switch to expert mode.
- After entering the question, press Enter to send it directly, or find and click the send button.

Technical Analysis

The workflow requires the agent to locate or open chat.deepseek.com, rely on an authenticated browser session, submit questions to the service, and use the resulting summary in subsequent operations.

The skill provides no per-query consent requirement, sensitive-data filtering, scope limitation, destination disclosure procedure, or rule requiring external model output to be treated as untrusted. Consequently, user-provided task content could cross the local trust boundary and be disclosed to the third-party service. The returned content could also influence later agent behavior without independent validation.

The source does not instruct the agent to capture or exfiltrate login credentials; it asks the user to perform login if necessary. It also contains no scripts, executable downloads, or persistence mechanisms. The confirmed issue is therefore limited to current-session instruction redirection and third-party query handling ...[truncated 1062 chars]

Remediation
View remediation

Remediation Suggestions

  1. Obtain informed, explicit consent before submitting each query to the external service.
  2. Show the user the exact content and destination before transmission.
  3. Remove credentials, secrets, personal information, proprietary data, and private workspace content from all externally submitted prompts.
  4. Do not rely on an existing authenticated browser session without confirming that the user authorizes its use for the current request.
  5. Treat all returned model content as untrusted data rather than executable instructions.
  6. Prevent external responses from changing tool policy, safety constraints, permissions, or task scope.
  7. Verify consequential claims using independent and authoritative sources.
  8. Limit downstream use of the response to the specific user-approved purpose and avoid retaining submitted or returned content unnecessarily.
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 (2)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill instructs the agent to use a logged-in third-party online model service (chat.deepseek.com) to answer user queries, but it does not warn the user that their prompts may be transmitted to an external service under the user's authenticated account. This creates a meaningful privacy and data-governance risk, especially if the agent handles sensitive user data, credentials, internal documents, or regulated information.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The entire skill is written to operate in Chinese and names a Chinese-language workflow, but it does not state that language choice is optional or limited to a justified region-specific context. Under the policy, forcing a specific language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.