Skills extract-from-pdfs

📦

extract-from-pdfs

Name: extract-from-pdfs
Author: brunoasm

Content revision r2 Critical 🌐 Network access🔑 Env variables📁 Filesystem access⚙️ External commands

Extract Structured Data from Scientific PDFs

Manual extraction from scientific literature is slow, inconsistent, and difficult to validate. This skill organizes papers, extracts schema-based records, validates values, and exports analysis-ready datasets.

Supports: Claude Codex Code(CC)

⚠️ 38 Poor

Do not auto-install this skill.

The canonical policy requires operator review before any installation action.

Install with my Agent

Copy this request to your Agent. It includes the canonical Skill page and manifest.

Agent request

Review the Skillstore skill "extract-from-pdfs" from https://skillstore.io/skills/brunoasm-extract-from-pdfs.md and its manifest at https://skillstore.io/api/skills/brunoasm-extract-from-pdfs/manifest. Verify the artifact. Do not auto-install. Inspect the skill and report your findings, then wait for an operator or manual installation decision.

Your Agent should still show its plan and request any confirmation required by the security policy.

Agent-readable resources

Use these links when an AI agent, crawler, or script needs clean context instead of reading the full page.

Markdown detail GET /skills/brunoasm-extract-from-pdfs.md Signed manifest GET /api/skills/brunoasm-extract-from-pdfs/manifest Signed lockfile GET /api/skills/brunoasm-extract-from-pdfs/lockfile

Test it

Using "extract-from-pdfs". A BibTeX library containing 146 entries and a directory of downloaded papers.

Expected outcome:

Organized 146 metadata records from the supplied BibTeX file.
Matched 132 records to PDFs and listed 14 missing files.
Saved the reviewed metadata catalog to metadata.json.

Using "extract-from-pdfs". Approved ecology PDFs and a schema for species, locations, methods, and measurements.

Expected outcome:

Processed 40 approved PDFs and produced structured records for 36 papers.
Flagged four failed extractions for review.
Preserved evidence notes and missing values according to the schema.

Using "extract-from-pdfs". A manually annotated validation sample and automated extraction results.

Expected outcome:

Calculated overall precision, recall, and F1.
Reported field-level scores and highlighted fields below the review threshold.
Exported sanitized analysis records and a separate validation report.

Security Audit

Critical

v8 • 7/23/2026 Open versioned report

The skill performs legitimate research processing, and most scanner hits are Markdown or ordinary local output operations. Confirmed risks include unchecked local-file uploads to Anthropic, a pipe-to-shell installer, plaintext shell-profile credential persistence, outbound scientific API requests, and unencrypted HTTP endpoints. Publication should wait until file-path validation, safer installation guidance, credential storage, transport security, and spreadsheet sanitization are addressed.

Files scanned

5,302

Lines analyzed

Review items

False positives ignored

Confirmed security concerns (5)

Critical

references/setup_guide.md:73

Pipe to shell pattern

curl -fsSL https://ollama.com/install.sh | sh

The guide pipes a remote installation script directly into sh without inspection, pinning, or checksum verification. A compromised response would execute arbitrary code immediately.

Critical

scripts/03_extract_from_pdfs.py:403-420 scripts/03_extract_from_pdfs.py:282-311

Unchecked metadata paths can upload arbitrary local files

The extraction pipeline trusts each metadata pdf_path, reads any existing local file without root, extension, symlink, or PDF-signature checks, then sends those bytes through Anthropic requests. Crafted metadata can therefore disclose files outside the intended PDF collection.

The code directly converts metadata-controlled paths to Path objects, reads all bytes after only exists(), and places the encoded bytes in Anthropic batch requests.

Medium

scripts/02_filter_abstracts.py:123-166 scripts/03_extract_from_pdfs.py:222-244

Untrusted document text can influence model decisions

Titles, abstracts, and PDF contents are placed in model user messages without an instruction hierarchy that treats embedded document commands as untrusted. Malicious paper text can alter filtering or extracted records, although the resulting model output has no direct tool privileges.

Both scripts submit untrusted research content as model input and trust structured model responses for workflow decisions. The impact is limited to data integrity because no tools execute from responses.

Medium

scripts/02_filter_abstracts.py:274-282 scripts/03_extract_from_pdfs.py:218-244

Research documents are transmitted to third-party model services

Anthropic filtering sends titles and abstracts, while PDF extraction uploads complete document bytes. This behavior is central to the skill but requires clear consent and handling rules for confidential, licensed, or sensitive literature.

The Anthropic client calls visibly include abstract text or base64-encoded PDF data. The code and skill documentation explicitly describe remote Claude processing.

Medium

scripts/06_export_database.py:127-142 scripts/06_export_database.py:242-255

Spreadsheet exports do not neutralize formulas

Extracted values are written directly to CSV and Excel without escaping leading formula characters. A malicious PDF or compromised model response can produce cells that spreadsheet software evaluates when a user opens the export.

Both export paths pass untrusted extracted strings directly to CSV or Excel writers, and no formula-prefix sanitization is present.

Capability review items (15)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

High

references/setup_guide.md:43

Hidden file in home directory

Add to your shell profile (~/.bashrc, ~/.zshrc) for persistence:

The guide recommends persisting credentials in executable shell profile files. This creates durable plaintext secret exposure and affects future shell sessions.

High

references/setup_guide.md:46

Hidden file in home directory

echo 'export ANTHROPIC_API_KEY="your-api-key-here"' >> ~/.bashrc

This command appends an Anthropic API key assignment to ~/.bashrc in plaintext. Other local processes or account users may read the persistent credential.

High

references/setup_guide.md:43

Hidden file access

Add to your shell profile (~/.bashrc, ~/.zshrc) for persistence:

The setup guidance directs users to modify hidden executable shell profiles for credential persistence. That is security-sensitive persistence, not ordinary project-file access.

High

references/setup_guide.md:46

Hidden file access

echo 'export ANTHROPIC_API_KEY="your-api-key-here"' >> ~/.bashrc

The command writes a plaintext API credential into ~/.bashrc. Persistent secret storage in a broadly readable shell profile creates a real credential-exposure risk.

High

references/setup_guide.md:46

Generic API/secret keys

echo 'export ANTHROPIC_API_KEY="your-api-key-here"' >> ~/.bashrc

The documented command persists the user's Anthropic API key in plaintext in ~/.bashrc. This is insecure credential handling even though the shown value is a placeholder.

Medium

scripts/01_organize_metadata.py:238

Python shutil operations

shutil.copy2(source_path, dest_path)

The copy operation uses a destination derived from an unsanitized metadata record ID. Crafted citation metadata can introduce path traversal and overwrite a PDF outside the intended output directory.

Medium

scripts/02_filter_abstracts.py:203

Python HTTP libraries

response = requests.post(

This is a real POST to the user-configurable Ollama URL and includes paper titles and abstracts. A remote or misconfigured endpoint receives research content.

Medium

scripts/05_validate_with_apis.py:111

Python HTTP libraries

response = requests.get(url, timeout=10)

The validator sends scientific names to World Flora Online over unencrypted HTTP. Network observers can read or alter the query and response.

Medium

scripts/05_validate_with_apis.py:151

Python HTTP libraries

response = requests.get(url, timeout=10)

The validator sends locations and the GeoNames username over unencrypted HTTP. This exposes account identifiers and research-derived location data to interception.

Medium

scripts/05_validate_with_apis.py:108

Hardcoded URL

url = f"http://www.worldfloraonline.org/api/1.0/search?query={quote(scientific_name)}"

The hardcoded World Flora Online endpoint uses plain HTTP for extracted scientific names. The lack of transport encryption creates confidentiality and integrity risk.

Medium

scripts/05_validate_with_apis.py:146

Hardcoded URL

url = f"http://api.geonames.org/searchJSON?q={quote(location)}&maxRows=1&username={username}"

The hardcoded GeoNames endpoint uses plain HTTP and places the username and location in the query string. Those values are exposed in transit.

Low

scripts/05_validate_with_apis.py:77

Python HTTP libraries

response = requests.get(url, timeout=10)

The validator sends extracted scientific names to the external GBIF service. The request is intentional, but it is genuine outbound disclosure of user-derived data.

Low

scripts/05_validate_with_apis.py:182

Python HTTP libraries

response = requests.get(url, headers=headers, timeout=10)

The geocoder sends extracted addresses to the external Nominatim service. HTTPS protects transit, but the request still discloses user-derived location data.

Low

scripts/05_validate_with_apis.py:213

Python HTTP libraries

response = requests.get(url, timeout=10)

The validator sends extracted compound names to PubChem. This is expected enrichment behavior but remains a real outbound data transfer.

Low

scripts/05_validate_with_apis.py:244

Python HTTP libraries

response = requests.get(search_url, timeout=10)

The validator sends gene symbols and optional organism names to NCBI. This is intended, but it is genuine external transmission of extracted values.

Risk Factors

🌐 Network access (30)

README.md:95 README.md:315 references/api_reference.md:273 references/api_reference.md:174 references/api_reference.md:200 references/api_reference.md:273 references/setup_guide.md:73 references/setup_guide.md:111 scripts/02_filter_abstracts.py:203 scripts/02_filter_abstracts.py:249 scripts/02_filter_abstracts.py:315 scripts/02_filter_abstracts.py:334 scripts/02_filter_abstracts.py:43 scripts/02_filter_abstracts.py:78 scripts/02_filter_abstracts.py:79 scripts/03_extract_from_pdfs.py:283 scripts/05_validate_with_apis.py:77 scripts/05_validate_with_apis.py:111 scripts/05_validate_with_apis.py:151 scripts/05_validate_with_apis.py:182 scripts/05_validate_with_apis.py:213 scripts/05_validate_with_apis.py:244 scripts/05_validate_with_apis.py:74 scripts/05_validate_with_apis.py:108 scripts/05_validate_with_apis.py:146 scripts/05_validate_with_apis.py:177 scripts/05_validate_with_apis.py:210 scripts/05_validate_with_apis.py:221 scripts/05_validate_with_apis.py:241 scripts/05_validate_with_apis.py:251

🔑 Env variables (18)

README.md:37 references/setup_guide.md:40 references/setup_guide.md:46 scripts/02_filter_abstracts.py:389 scripts/02_filter_abstracts.py:389 scripts/02_filter_abstracts.py:389 scripts/02_filter_abstracts.py:390 scripts/03_extract_from_pdfs.py:372 scripts/03_extract_from_pdfs.py:372 scripts/03_extract_from_pdfs.py:372 scripts/03_extract_from_pdfs.py:373 scripts/05_validate_with_apis.py:141 scripts/05_validate_with_apis.py:141 scripts/05_validate_with_apis.py:314 scripts/05_validate_with_apis.py:317 scripts/05_validate_with_apis.py:318 scripts/05_validate_with_apis.py:319 SKILL.md:47

📁 Filesystem access (20)

⚙️ External commands (48)

SKILL.md:41-43 SKILL.md:43-46 SKILL.md:46-47 SKILL.md:47-64 SKILL.md:64-66 SKILL.md:66-69 SKILL.md:69-70 SKILL.md:70-71 SKILL.md:71-72 SKILL.md:72-74 SKILL.md:74-82 SKILL.md:82-122 SKILL.md:122-128 SKILL.md:128-145 SKILL.md:145-151 SKILL.md:151-154 SKILL.md:154-156 SKILL.md:156-159 SKILL.md:159-161 SKILL.md:161-164 SKILL.md:164-166 SKILL.md:166-169 SKILL.md:169-171 SKILL.md:171-177 SKILL.md:177-189 SKILL.md:189-192 SKILL.md:192-193 SKILL.md:193-194 SKILL.md:194-195 SKILL.md:195-196 SKILL.md:196-197 SKILL.md:197-199 SKILL.md:199-203 SKILL.md:203-227 SKILL.md:227-228 SKILL.md:228-259 SKILL.md:259-264 SKILL.md:264-267 SKILL.md:267-270 SKILL.md:270-273 SKILL.md:273-274 SKILL.md:274-275 SKILL.md:275-278 SKILL.md:278-279 SKILL.md:279-284 SKILL.md:284-285 SKILL.md:285-288 SKILL.md:288-289

Detected Patterns

Pipe to shell pattern

Audited by: codex View Audit History →

Share & cite this report

Share the versioned assessment report, neutral badge, embed card, and citations. Skillstore reports evidence without deciding whether this Skill is safe.

Open versioned report

Copy report link

https://skillstore.io/skills/brunoasm-extract-from-pdfs/audits/8?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdown badge

[![Skillstore security assessment](https://skillstore.io/badges/skills/brunoasm-extract-from-pdfs/security.svg)](https://skillstore.io/skills/brunoasm-extract-from-pdfs?utm_source=security_passport_badge)

HTML badge

<a href="https://skillstore.io/skills/brunoasm-extract-from-pdfs?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/brunoasm-extract-from-pdfs/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

Embed card

<iframe src="https://skillstore.io/embed/skills/brunoasm-extract-from-pdfs.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

Academic citations (APA · BibTeX · CFF)

APA citation

brunoasm. (2026). extract-from-pdfs security audit report (audit version 8) [Author version unspecified]. Skillstore. https://skillstore.io/skills/brunoasm-extract-from-pdfs/audits/8

BibTeX citation

@techreport{brunoasm-brunoasm-extract-from-pdfs-2026,
  author = {brunoasm},
  title = {extract-from-pdfs security audit report (audit version 8)},
  institution = {Skillstore},
  year = {2026},
  number = {8},
  url = {https://skillstore.io/skills/brunoasm-extract-from-pdfs/audits/8},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "extract-from-pdfs security audit report (audit version 8)"
version: "unspecified"
type: report
authors:
  - name: "brunoasm"
date-released: "2026-07-23"
url: "https://skillstore.io/skills/brunoasm-extract-from-pdfs/audits/8"
identifiers:
  - type: other
    value: "skillstore:brunoasm-extract-from-pdfs:audit:8"
    description: "Skillstore immutable audit report identifier"

Skillstore Score

Why this score Evidence Confidence: Medium

Architecture

Maintainability

Content

Community

Spec Compliance

What You Can Build

Build a systematic review dataset

Convert a screened scientific literature collection into consistent records for evidence synthesis and meta-analysis.

Standardize a research archive

Organize mixed metadata sources, match PDFs, validate selected fields, and export a reusable database.

Measure extraction quality

Create annotation samples and compare automated records with ground truth using field-level precision, recall, and F1.

Try These Prompts

Organize a PDF collection

Organize the PDFs in [folder] using [BibTeX, RIS, directory, or DOI list]. Create metadata.json and report missing PDF links. Do not upload files.

Design an extraction schema

Review [example PDFs] and design a schema for extracting [research variables]. Include field types, required fields, evidence quotes, missing-value rules, and one example.

Run and validate extraction

Filter [metadata file] for [criteria], extract records from approved PDFs, repair outputs, and validate against [schema]. Use test mode before full processing.

Evaluate and export a review dataset

Create a stratified validation set from [results file]. Calculate field-level precision, recall, and F1, then export sanitized flattened records to [format].

Best Practices

Resolve and review every PDF path before any external upload.
Start with representative PDFs and refine the schema using explicit field types, evidence requirements, and missing-value rules.
Run small tests, annotate a validation sample, and inspect field-level metrics before processing the full collection.

Avoid

Do not run pipe-to-shell installers or store API keys in shell profiles.
Do not send confidential or unlicensed PDFs to external models without authorization.
Do not treat model output as ground truth or open unsanitized spreadsheet exports.