recoup-internal-upwork

Warn

Audited by Gen Agent Trust Hub on Aug 31, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to install the unslop plugin from https://www.skills.sh/cursor/plugins/unslop using npx skills add. This source is not verified. Additionally, it relies on a third-party Apify actor neatrat/upwork-job-scraper to fetch job data.
  • [REMOTE_CODE_EXECUTION]: The skill executes remote code through the installation of external plugins and the invocation of the Apify scraper via shell commands (curl piped to python3). These processes handle remote data from unverified sources.
  • [COMMAND_EXECUTION]: Several shell commands are provided for runtime execution, including:
  • curl and python3 for interacting with the Apify API using the $APIFY_TOKEN environment variable.
  • git commands for auditing local repository history (git log).
  • Google Chrome in headless mode for rendering HTML templates to PDF files.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from Upwork job postings and scraped results. This data is used to populate internal workspace files and generate cover letters. Malicious instructions embedded in a job posting could potentially influence the agent's behavior, leading to unauthorized actions or data exfiltration.
  • Ingestion points: Pasted job postings in JOB.md and scraped JSON data from the Apify actor.
  • Boundary markers: None identified; untrusted content is used directly in triage logic and requirement tables.
  • Capability inventory: The agent has capabilities for shell execution (curl, python3, git, chrome) and filesystem read/write access.
  • Sanitization: There is no evidence of sanitization or filtering for the external content before it is processed.
  • [DATA_EXPOSURE]: The skill is designed to programmatically read and analyze local Git repositories to verify professional claims. While this is the intended purpose, it requires the agent to have read access to source code on the local filesystem, which could lead to the exposure of sensitive information if the agent is compromised via indirect prompt injection.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 31, 2026, 07:05 PM
Security Audit — agent-trust-hub — recoup-internal-upwork