The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from the target AWS environment, including resource names, tags, and configuration metadata. This content is then summarized and analyzed by the agent. A malicious actor with access to the AWS environment could potentially craft resource metadata (e.g., a Security Group name or S3 bucket tag) containing prompt injection payloads to influence the final compliance report's conclusions.\n
Ingestion points: Reference files in references/programmatic-checks/ execute AWS CLI commands whose output is processed by the agent.\n
Boundary markers: While the skill has a structured phase-based workflow, it does not use explicit boundary delimiters for data ingested from the environment.\n
Capability inventory: The skill possesses Bash, Read, and Write capabilities across all files.\n
Sanitization: The skill uses standard utilities like awk and grep for parsing, but lacks specific sanitization for LLM-based injection attacks.\n- [DYNAMIC_EXECUTION]: The skill uses a python3 here-doc in references/programmatic-checks/ia-identification-authentication.md to process the IAM credential report. This involves generating and executing a script at runtime to calculate credential ages and rotation status.\n- [COMMAND_EXECUTION]: The skill relies on extensive execution of AWS CLI commands via Bash to perform its discovery and analysis phases. This is the primary function of the skill, and risks are mitigated by the mandatory read-only credential check in Phase 1.\n- [DATA_EXFILTRATION]: The skill accesses sensitive environment metadata, including IAM credential reports, password policies, and network configurations. This data is stored in local temporary files (e.g., /tmp/cred-report.csv) and formatted into the final report. While no external network exfiltration to unauthorized domains was detected, the skill handles highly sensitive security information.

cjis-reviewer