Back to skill

Security audit

deep-search-engine

Security checks for vulnerabilities and agentic risk

Overview

The skill is rated suspicious because scanners found privacy and reliability weaknesses in its search and verification workflow, with incomplete coverage of some references; they did not report clear malicious behavior.

Treat this as a review-level warning, not a finding of malicious intent. Before installing, consider whether your searches might use confidential or unpublished material and whether external disclosure is acceptable. Search and verification results may be unreliable: dates and fingerprints can be accepted from unrelated page content, and PASS may be returned without date evidence. Skillspector inspected all seven listed components, but coverage is marked partial because three path-like references could not be resolved; the AIG scan completed with no findings.

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
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (6)

Code capability 'network' detected in scripts/baidu_search.py but not covered by declared permissions.

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'network' capability that is not covered by the declared 'allowed-tools' preapproval. Review the host runtime's actual enforcement policy; runtime permissions are unknown unless separately verified, and this declaration alone does not establish a permission ceiling.

Content

No source excerpt is available for this finding.

The workflow sends excerpts of user-provided source content to third-party search services without a privacy warning.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow extracts distinctive sentences from user-provided material and submits them to external search providers without checking whether that material is public or authorized for disclosure. Publication monitoring makes such queries legitimate for public articles, but unpublished drafts or confidential campaign material could be disclosed through search queries and provider logs.

Content

No source excerpt is available for this finding.

Arbitrary dates in the page are treated as evidence that the publication date falls within the requested window.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Date extraction collects matching strings throughout the HTML, and window matching accepts any of them without establishing that they describe publication. A sidebar, event reference, or attacker-controlled hidden date can therefore make an out-of-window article appear verified; mandatory manual review mitigates this but does not correct the script's verdict.

Content

No source excerpt is available for this finding.

Fingerprint verification accepts matches anywhere in raw HTML rather than requiring article-body evidence.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The fingerprint condition accepts matches anywhere in raw HTML, including scripts, comments, metadata, and hidden elements, rather than requiring evidence in the article's visible content. An adversarial page can plant the expected text to obtain a verification signal without publishing it as article content; the documented manual check reduces exposure but leaves the automated result unreliable.

Content

No source excerpt is available for this finding.

PASS can be returned without any publication-date evidence, contradicting the documented three-element requirement.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

When --window is omitted, the PASS condition explicitly allows a missing date_hit, so an undated page can satisfy a purported three-element verification. The skill's mandatory window and manual-review instructions reduce normal-workflow exposure, but the script still emits a misleading machine-readable PASS that direct callers or downstream automation can trust.

Content

No source excerpt is available for this finding.

The cleaner deletes potentially functional parameters despite documenting that all remaining parameters are preserved.

Low
Category
Excessive Agency
Confidence
98% confidence
Finding

The cleaner drops valueless parameters and parameters with values such as 0 regardless of their meaning, potentially changing resource selection or page behavior. This can corrupt evidence links in a verification report, although the skill explicitly requires checking content equivalence and retaining the original URL when uncertain.

Content

No source excerpt is available for this finding.