auto-paper-improvement-loop

Pass

Audited by Gen Agent Trust Hub on Nov 1, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the user's paper (LaTeX source files and PDFs) as well as external style references (from URLs, arXiv IDs, or local files). This content is concatenated and passed directly to the gpt-6-astra model for review. Because the model is instructed to follow complex review guidelines, an attacker could embed malicious instructions within a paper or style reference to influence the agent's behavior.
  • Ingestion points: Step 1 (collecting paper text) and the --style-ref parameter (fetching external sources).
  • Boundary markers: The prompt uses simple markdown headers (e.g., ## Paper Files:) to separate content from instructions, which can be bypassed by sophisticated injections.
  • Capability inventory: The agent possesses broad capabilities including shell execution (Bash), file modification (Edit, Write), and the ability to call other skills.
  • Sanitization: No sanitization is applied to the ingested LaTeX content or external style data.
  • [DATA_EXFILTRATION]: The skill accesses sensitive local data by reading ~/.claude/feishu.json to send notifications to the Feishu platform. It also transmits the full text and PDF content of the paper being improved to an external LLM via the mcp__codex__codex tool. While these operations are functional, they involve the external transmission of potentially sensitive researcher data and service credentials.
  • [COMMAND_EXECUTION]: The skill relies heavily on shell command execution for its core functionality, including paper compilation (latexmk), document parsing (grep, awk), and structured data manipulation (jq). It also interacts with other local skill primitives such as /kill-argument.
  • [DYNAMIC_EXECUTION]: The skill employs dynamic path resolution to find and execute the extract_paper_style.py helper script based on the local environment and the presence of ARIS repository configuration. Additionally, it executes an inline Python script via a shell heredoc in Step 4.5 to perform theorem normalization and consistency checks.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 1, 2026, 01:42 AM