The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides a fallback installation method that uses 'curl -fsSL https://cohesivity.ai/quickstart.sh | bash'. This pattern downloads and executes a remote script in the system shell without any integrity checks or human review, which is a critical vulnerability for malicious script injection.
[COMMAND_EXECUTION]: The skill instructs the agent to run 'npx @cohesivity/init' as its primary bootstrap method, granting the skill the ability to execute code and install Node.js packages on the local system.
[EXTERNAL_DOWNLOADS]: The skill performs multiple network operations to 'cohesivity.ai', including fetching live documentation from 'https://cohesivity.ai/offerings/', pricing data, and its own latest version. This creates a reliance on external infrastructure that could be compromised.
[CREDENTIALS_UNSAFE]: The skill stores highly sensitive 'coh_management_key' and 'coh_application_key' in a local '.cohesivity' file within the project directory. While it advises against client-side exposure, the storage of cleartext management keys on disk is a significant data exposure risk.
[INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface for indirect prompt injection. 1. Ingestion points: It dynamically fetches and processes content from 'https://cohesivity.ai/offerings/' and 'llms.txt'. 2. Boundary markers: The skill does not define boundary markers to prevent the agent from following instructions found in these external documents. 3. Capability inventory: The skill possesses shell execution capabilities, file system write access, and network access. 4. Sanitization: There is no evidence of sanitization or validation of the fetched remote documentation before it is integrated into the agent's operational context.

cohesivity