The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse untrusted data from various external M&A artifacts, creating a surface for indirect prompt injection attacks.
Ingestion points: The skill reads deal-context.md, closing-checklist.yaml, and user-uploaded documents such as Equity Purchase Agreements (PA), disclosure schedules, and contract lists. It also attempts to load files from connected repositories (e.g., Feishu, cloud documents).
Boundary markers: There are no explicit instructions or delimiters defined to separate the data content from the agent's instructions, nor are there warnings to ignore embedded commands within the ingested legal text.
Capability inventory: The skill has the capability to write to local configuration files (~/.claude/plugins/config/...) and export data in CSV/Table formats.
Sanitization: While the skill includes a specific "Formula Injection Defense" to escape characters like =, +, and @ in exported CSV files, it lacks sanitization or validation for natural language instructions that might be embedded in the source documents to manipulate the agent's logic.
[DATA_EXFILTRATION]: The skill interacts with highly sensitive corporate data and local file paths.
Sensitive paths: The skill reads and writes to ~/.claude/plugins/config/claude-for-legal-zh/corporate-legal/deals/[代码]/integration-tracker.yaml. While specific to the plugin, these paths contain summarized M&A transaction data.
Sensitive data exposure: The agent is instructed to extract and process financial milestones, escrow release dates, and intellectual property transfer details from purchase agreements, which are highly confidential.

integration-management