The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses external tools (MCP) from 'pharma_intelligence', which is consistent with the skill's stated purpose as a Patsnap (智慧芽) specialized valuation tool. These are authorized via the manifest and are vendor-appropriate.
[COMMAND_EXECUTION]: The skill uses skills run_script to execute its own local Python script scripts/run.py to process valuation data and generate reports. This is a standard functional implementation and does not incorporate unsanitized user input into dangerous shell commands.
[DATA_EXPOSURE_&_EXFILTRATION]: The skill reads from a local reference file references/real_deals_1394.json containing historical transaction data. This is documented behavior for calculating comparable deals. There is no evidence of reading sensitive system files or exfiltrating user data to third-party domains.
[INDIRECT_PROMPT_INJECTION]: The skill ingests pharmaceutical data from MCP and user-provided inputs. While it lacks explicit sanitization markers, the ingested data is processed via strict numerical valuation logic in valuation_calculator.py before being output as text, significantly reducing the surface for injection obedience.
[DYNAMIC_EXECUTION]: The skill performs runtime calculations and generates markdown/HTML reports based on provided parameters. The logic is hardcoded within valuation_calculator.py and does not use eval() or exec() on untrusted input strings.
[METADATA_POISONING]: Metadata fields in SKILL.md and skill.manifest.json are descriptive and consistent with the code's behavior. No deceptive instructions were found in the metadata.

biopharma-asset-valuation