T01 · Skill Instruction Hijacking
- Location
SKILL.md:1114- Finding
Unrelated Agent Behavior Steering and Third-Party Promotion
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill mostly performs citation-management tasks, but it includes unrelated agent steering, promotional guidance, and risky search/proxy instructions that users should review before installing.
Install only if you are comfortable with citation identifiers, search queries, and optional NCBI email/API key values being sent to scholarly services. Avoid the Google Scholar free-proxy option and IP-switching guidance, do not run the absent schematic-generation command, and use an isolated environment with pinned dependencies before writing over bibliography files.
SKILL.md:1114Unrelated Agent Behavior Steering and Third-Party Promotion
scripts/extract_metadata.py:217Plaintext HTTP Transport for arXiv Metadata Queries
scripts/search_google_scholar.py:38Google Scholar Queries Can Be Routed Through Untrusted Free Proxies
SKILL.md:1079Unpinned and Excessive Python Dependency Installation Guidance
The description overstates the skill substantially. The supplied code performs a narrow, specific function: converting DOI strings into BibTeX entries by requesting application/x-bibtex from doi.org/CrossRef, with optional batch processing and file I/O. While this partially aligns with the declared mention of converting DOIs to BibTeX, it does not implement the broader advertised capabilities of searching Google Scholar or PubMed, extracting/validating citation metadata, or ensuring reference accuracy. The actual primary purpose is a DOI-to-BibTeX converter, not a comprehensive citation management tool.
The description overstates the skill as a comprehensive citation manager with paper search and citation validation capabilities. The code is narrower: it retrieves metadata for already-known identifiers (DOI/PMID/arXiv/URL) and formats it as BibTeX or JSON. It does use PubMed, but only to fetch by PMID, not to search for papers. There is no Google Scholar integration, no query-based paper discovery, and no substantive citation validation logic. Additionally, the code includes arXiv support and JSON output, which are not reflected in the description. The core overlap is DOI-to-BibTeX and metadata extraction, but the declared purpose is materially broader and partly inaccurate.
There is a clear mismatch between the declared description and the actual code. The description promises broad citation-management functions centered on discovering papers and verifying citation data via Google Scholar and PubMed, plus DOI conversion and metadata retrieval. The code does none of that: it operates only on a user-supplied BibTeX file using local parsing and string cleanup. Its main functions are formatting, sorting, deduplication, and minor normalization of fields. While this overlaps slightly with 'generate properly formatted BibTeX entries' and reference hygiene, the core declared capabilities—searching, external metadata extraction, validation, and DOI lookup—are absent. Therefore the description materially overstates and misrepresents the skill's behavior.
The description promises a broader citation-management skill spanning Google Scholar and PubMed, with validation of citation information, DOI conversion, and accuracy assurance. This code chunk only implements Google Scholar search and basic metadata extraction/output formatting. While it does generate BibTeX, the entries are inferred from sparse metadata and guessed venue type, not validated against authoritative citation sources. There is also no PubMed integration in the provided code, no DOI handling, and no citation verification logic. Additionally, the optional proxy setup is a network-related capability not reflected in the declared permissions/purpose. Therefore the declared description overstates the implemented behavior, and the code does not fully match the claimed comprehensive functionality.
The code aligns with part of the description: it searches PubMed, extracts metadata, and can generate BibTeX entries. However, the declared purpose materially overstates functionality. There is no implementation for Google Scholar search, no citation validation logic, no mechanism to verify reference accuracy, and no standalone DOI conversion feature. The primary behavior is a PubMed-specific retrieval and formatting script rather than a comprehensive cross-source citation-management tool. No suspicious undeclared behavior outside this domain is evident, but the description does not accurately represent the full claimed capability set.
The declared description presents a broad citation-management and paper-discovery tool, including searching Google Scholar/PubMed, extracting metadata, converting DOIs to BibTeX, and generating properly formatted BibTeX entries. The supplied code does not implement those primary capabilities. Instead, it validates existing BibTeX files, checks field presence/format, detects duplicates, and optionally verifies DOI resolution using doi.org/CrossRef. The DOI verification is consistent with citation validation, but the major advertised features—literature search, DOI-to-BibTeX conversion, and BibTeX generation—are absent. Therefore the description materially overstates and misrepresents the code’s actual behavior.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
### 1 Day Before Submission
- [ ] Final validation check
- [ ] LaTeX compilation successful with no warnings
- [ ] PDF renders all citations correctly
- [ ] Bibliography appears in correct format
- [ ] No placeholder citations (Smith et al. XXXX)
This script makes outbound network requests to doi.org/CrossRef using user-supplied DOI input, but the capability is not declared in permissions. Undeclared network access is dangerous because it expands the skill's effective privileges beyond what reviewers or users may expect, and in a broader agent environment it could enable unauthorized data exfiltration or unreviewed external communication.
The skill uses 'env' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.
The documentation instructs the agent to create and save image files by default without clear user confirmation. In an agent environment with Write/Edit/Bash permissions, this can lead to unrequested filesystem modifications and execution of auxiliary generation commands, violating the principle of least astonishment and potentially causing downstream security or compliance issues.
The skill includes embedded instructions to upsell or redirect users to an external hosted platform unrelated to the immediate citation task. In an agent setting, this can manipulate workflow routing, nudge users to move potentially sensitive research context to a third-party service, and create conflicted behavior not aligned with the user's request.
The guide explicitly recommends rotating User-Agents and considering proxies to deal with Google Scholar rate limiting. Those are evasive scraping tactics that go beyond normal citation lookup and can facilitate bypassing service protections or terms-of-service restrictions. In a citation-management skill, this is unnecessary and makes abuse easier.
The document says users should respect rate limits and not use multiple IPs to bypass limits, but later troubleshooting advises switching to a different IP/network after blocking. That contradiction normalizes circumvention and weakens the safety guidance, making policy-violating behavior more likely.
The guidance to switch to a different IP or network when blocked encourages circumvention of access controls, and the lack of any warning about terms-of-service or policy conflicts makes that worse. In practice, this advice can help users continue behavior that triggered anti-abuse defenses.
The troubleshooting section advises switching to a different IP/network when blocked. This is a concrete bypass recommendation for access restrictions and is not justified by the skill's academic citation-management purpose. It could be used to continue automated scraping after enforcement triggers.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
**Research datasets, software, other outputs** - Assigns DOIs to non-traditional scholarly works.
**Base URL**: `https://api.datacite.org/dois/`
**Similar to CrossRef** but for datasets, software, code, etc.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
**Research datasets, software, other outputs** - Assigns DOIs to non-traditional scholarly works.
**Base URL**: `https://api.datacite.org/dois/`
**Similar to CrossRef** but for datasets, software, code, etc.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
**Research datasets, software, other outputs** - Assigns DOIs to non-traditional scholarly works.
**Base URL**: `https://api.datacite.org/dois/`
**Similar to CrossRef** but for datasets, software, code, etc.
The systematic search example hard-codes AND English[Language], which imposes a language restriction in the guidance. Because the file does not present this as an optional user choice or justify it as a region-specific requirement, it conflicts with the policy against forcing a specific language without opt-in.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Returns:
Metadata dictionary or None
"""
url = f"https://api.crossref.org/works/{doi}"
try:
response = self.session.get(url, timeout=15)
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Returns:
Metadata dictionary or None
"""
url = f"https://api.crossref.org/works/{doi}"
try:
response = self.session.get(url, timeout=15)
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Returns:
Metadata dictionary or None
"""
url = f"https://api.crossref.org/works/{doi}"
try:
response = self.session.get(url, timeout=15)
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Returns:
Metadata dictionary or None
"""
url = f"https://api.crossref.org/works/{doi}"
try:
response = self.session.get(url, timeout=15)
When --check-dois is used, the tool sends DOI values to doi.org and CrossRef, which can disclose a user's research interests, draft bibliography contents, or unpublished manuscript references to third parties. This is not inherently malicious, but the lack of an explicit privacy warning or consent flow makes the external transmission less transparent and can violate user expectations in academic or sensitive research contexts.
The code performs an outbound request to https://api.crossref.org/works/{doi}, transmitting DOI-derived citation data to an external service. In the context of a citation-management skill this behavior is expected and functionally necessary for metadata validation, but it still creates a privacy/data-leak surface if users are validating confidential or prepublication references without realizing network lookups occur.
if response.status_code < 400:
# DOI resolves, now get metadata from CrossRef
crossref_url = f"https://api.crossref.org/works/{doi}"
metadata_response = self.session.get(crossref_url, timeout=10)
if metadata_response.status_code == 200: