azure
Fail
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [CREDENTIALS_UNSAFE]: Hardcoded default password 'Contoso!Passw0rd2024' found in the example Bicep template located at 'evals/files/main.bicep'.
- [DATA_EXFILTRATION]: The provisioning hook in 'evals/files/azure.yaml' executes a command to extract a Cosmos DB primary master key and store it in cleartext application settings, exposing sensitive data to the management plane.
- [COMMAND_EXECUTION]: Shell scripts in 'evals/files/provision-aks.sh' and 'evals/files/azure.yaml' perform automated infrastructure management using 'az' and 'kubectl' commands, including credential handling and cluster configuration.
- [EXTERNAL_DOWNLOADS]: The skill fetches real-time pricing information from Microsoft's official Retail Prices API (prices.azure.com) in 'references/cost-and-quotas.md'.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided Bicep and YAML infrastructure files using powerful CLI tools without explicit boundary markers, potentially allowing malicious code in reviewed files to influence agent actions. Ingestion points: Files like '.bicep', 'azure.yaml', and shell scripts are read for review in 'SKILL.md'. Boundary markers: Absent; there are no instructions to ignore instructions inside reviewed files. Capability inventory: Resource management, credential fetching, and shell execution via Azure CLI and 'kubectl'. Sanitization: None detected.
Recommendations
- AI detected serious security threats
Audit Metadata