parallel-web
Audited by Socket on Feb 22, 2026
1 alert found:
Malware[Skill Scanner] Natural language instruction to download and install from URL detected All findings: [CRITICAL] command_injection: Natural language instruction to download and install from URL detected (CI009) [AITech 9.1.4] [CRITICAL] command_injection: Natural language instruction to download and install from URL detected (CI009) [AITech 9.1.4] [HIGH] supply_chain: Installation of third-party script detected (SC006) [AITech 9.1.4] [HIGH] supply_chain: Installation of third-party script detected (SC006) [AITech 9.1.4] This skill is functionally consistent with its stated purpose (web search, deep research, URL extraction) but contains multiple supply-chain and operational risk signals: it requires installing an unverified package ('parallel-web') without pinning, mandates persistent local storage of all fetched content, and grants the agent shell-level capabilities (Bash) which broaden what the agent can do with fetched data and installed packages. There is no direct evidence of malware or covert exfiltration in the documentation alone, but the combination of third-party install, unpinned dependency, mandatory write-to-disk, and broad allowed-tools raises the chance of misuse or unintended data exposure. I rate this as medium security risk (not confirmed malware). Recommend: verify the provenance of the 'parallel-web' package, pin package versions and checksums, minimize allowed-tools (remove Bash if possible), and make saving to disk an opt-in behavior with clear consent and sanitization policies. LLM verification: The skill's stated purpose and capabilities are internally consistent: it legitimately requires an API key and network access to Parallel.ai, and saving results locally fits the stated reproducibility goals. The main security concerns are unpinned pip dependencies (supply-chain risk) and the mandatory policy to save all results to disk (could persist sensitive data). There is no evidence in the provided text of obfuscated code, hardcoded secrets, credential harvesting, or explicit exfiltration t