aws-iam
Warn
Audited by Gen Agent Trust Hub on Jul 24, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill directs the agent to search for AWS environment information within potentially sensitive project files.
- Instructions in
references/aws-iam-policy-generation.md(Task 1b) mandate checking.env,terraform.tfvars,cdk.json,cdk.context.json, andsamconfig.tomlfor account IDs and regions. Accessing these files, which are standard locations for storing secrets, poses a significant data exposure risk. - [REMOTE_CODE_EXECUTION]: The skill uses
uvxto download and execute an external tool for analyzing source code at runtime. - Executes
uvx iam-policy-autopilot@latest generate-policiesto derive IAM policies from application code. The tool is maintained by theawslabsorganization on GitHub, an official repository for Amazon Web Services. - [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to interact with the environment and development tools.
- Runs
aws sts get-caller-identityto retrieve current account details. - Executes the
iam-policy-autopilotCLI to generate permissions. - [EXTERNAL_DOWNLOADS]: The skill fetches official AWS service definitions from a remote endpoint.
- Queries
https://servicereference.us-east-1.amazonaws.comto map API operations to the correct IAM actions. - [PROMPT_INJECTION]: The skill ingests untrusted application source code to automate policy generation, creating a surface for indirect prompt injection.
- Ingestion points: Analyzes source code files in Python, Go, TypeScript, and Java provided by the user in
references/aws-iam-policy-generation.md. - Boundary markers: No specific delimiters or safety instructions are defined for the source code content being processed.
- Capability inventory: The skill has access to shell execution (
aws,uvx) and network operations (curl). - Sanitization: Risk is mitigated by the use of the
iam-policy-autopilottool, which performs deterministic static analysis rather than relying on LLM reasoning to interpret the code, reducing the likelihood of malicious instructions influencing the output.
Audit Metadata