Back to skill

Security audit

Academic Search

Security checks for vulnerabilities and agentic risk

Overview

This academic search skill is purpose-aligned and mainly instructs the agent to query scholarly sources, with some routing and transport-quality issues users should understand.

Install only if you are comfortable with academic search queries being sent to external scholarly services. The publisher should narrow the broad triggers and prefer HTTPS for arXiv links/endpoints, but the inspected artifacts do not show deception, persistence, credential access, or destructive behavior.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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)

T09 · Insecure Skill Coding Practices

Note
Location
knowledge/domain.md:13
Finding

Unencrypted arXiv API Transport Permits Search Result Manipulation

Content
View full analysis

Vulnerability Details

File Location: knowledge/domain.md, lines 13-18
Vulnerability Type: Use of an unencrypted HTTP endpoint
Risk Level: Low

Vulnerable Code

markdown
### API Endpoint
- **Base URL**: `http://export.arxiv.org/api/query`
- **Method**: GET
- **Rate Limit**: 1 request per 3 seconds (be respectful; no authentication required)

Technical Analysis

The Skill explicitly instructs the Agent to query the arXiv API through cleartext HTTP. HTTP does not provide transport confidentiality, server authentication, or response integrity.

An attacker with a network-level interception position could observe academic search terms or modify the returned Atom feed. Modified fields could include paper titles, authors, abstracts, identifiers, and links. Because the workflow subsequently screens, ranks, and summarizes this metadata, manipulated records could enter the final literature review.

Cleartext arXiv links also appear in sample response data at knowledge/domain.md:63 and knowledge/domain.md:71. Those examples reinforce the use of insecure transport but are not the primary API configuration.

Attack Path

  1. The Agent follows the documented endpoint and submits an arXiv query over HTTP.
  2. An attacker capable of intercepting network traffic observes the request.
  3. The attacker alters the HTTP response before it reaches the Agent.
  4. The modified metadata is processed by the Skill's screening, citation-analysis, and ranking workflow.
  5. The Agent may present inaccurate citations, fabricated paper metadata, or substituted links to the user.

Impact Assessment

The issue does not grant local system privileges, code execution, persistent access, or credential access. The affected scope is limited primarily to the confidentiality of academic search queries and the integrity of arXiv search results.

Successful exploitation could undermine bibliographic accuracy, contaminate ...[truncated 71 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace the cleartext endpoint with HTTPS:

    markdown
    - **Base URL**: `https://export.arxiv.org/api/query`
    
  2. Update the example response links at knowledge/domain.md:63 and knowledge/domain.md:71 to use HTTPS.

  3. Require HTTPS for all academic database requests and retrieved document links.

  4. Reject or warn on redirects that downgrade a request from HTTPS to HTTP.

  5. Validate returned identifiers and links against expected arXiv formats and trusted hostnames before including them in results.

  6. Cross-check important bibliographic metadata against an independent source, such as Semantic Scholar, before presenting it as verified.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (10)

Shadow Command Trigger

Medium
Category
Trigger Abuse
Confidence
84% confidence
Finding

The trigger phrase 'find papers' overlaps with a common built-in verb pattern centered on 'find', creating a shadow-trigger condition where this skill may intercept requests meant for more general search or retrieval behavior. In agent systems, command ambiguity can route user input to the wrong skill, causing unintended tool use, confusing outputs, or bypass of expected controls for other handlers.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger list contains very broad terms such as 'research', 'scholar', 'cite', and 'citation' that can match many ordinary user requests outside the intended academic-paper-search workflow. This can cause unintended skill activation, leading the agent to invoke specialized instructions or external-search behavior in contexts where the user did not explicitly request it.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · knowledge/anti-patterns.md (reported line 73)May include surrounding context.

md
- **Fix**: Only return papers actually retrieved from database API responses. If the search returns fewer than 5 relevant results, report that honestly rather than padding with fabricated entries. Include the database query used so the user can verify.

### 15. Ignoring Open Access Status
- **Problem**: Returning papers behind paywalls without checking for open-access alternatives. Users without institutional access cannot read the papers.
- **Fix**: For each paper, check: (a) arXiv preprint version, (b) Semantic Scholar `openAccessPdf` field, (c) author's personal website or institutional repository. Flag papers with no open-access version available and suggest the user check their institutional library access.

## Workflow Anti-Patterns

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 95)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 109)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 113)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 114)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 123)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · knowledge/domain.md (reported line 124)May include surrounding context.

md
### API Endpoints

#### Paper Search
- **URL**: `GET https://api.semanticscholar.org/graph/v1/paper/search`
- **Rate Limit**: 100 requests/5 minutes (unauthenticated); 1 request/second with API key

| Parameter | Description | Example |

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger list includes very broad terms such as "research," "arxiv," "scholar," "cite," and "citation," which can appear in many normal conversations and may cause the skill to activate outside the user's intended scope. In an agentic system, unintended activation can route queries to this skill when another tool or no tool would be more appropriate, creating prompt-routing errors, unnecessary external lookups, and increased exposure to downstream tool misuse.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.