huawei-cloud-cce-chaos-experiment
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from CCE container logs and Kubernetes events (scripts/collect_logs.py). This content is parsed for error patterns and then interpolated into Markdown reports (scripts/generate_analysis_report.py), creating a potential surface for indirect prompt injection if log content contains adversarial instructions.
- Ingestion points: scripts/collect_logs.py and scripts/analyze_logs.py ingest logs and event data from the cluster.
- Boundary markers: No specific delimiters or warnings for the agent to ignore embedded instructions are present.
- Capability inventory: scripts/execute_experiment.py performs ECS BatchStopServers (shutdown); scripts/rollback_experiment.py performs ECS BatchStartServers (startup).
- Sanitization: No sanitization or escaping of external content is performed; the analysis relies on simple keyword matching.
- [DYNAMIC_EXECUTION]: scripts/deploy_experiment.sh dynamically generates the rollback_experiment.sh shell script at runtime using heredoc templates and local variables to facilitate emergency recovery operations.
- [COMMAND_EXECUTION]: Multiple Python and Bash scripts, including scripts/execute_experiment.py and scripts/discover_cce.py, utilize subprocess calls to execute hcloud and kubectl commands for managing cloud infrastructure and cluster state.
- [EXTERNAL_DOWNLOADS]: scripts/check_env.sh fetches the kubectl binary from official Kubernetes distribution sites and Huawei Cloud mirrors. Documentation also references the installation of the hcloud CLI tool from a vendor-owned Huawei Cloud OBS bucket.
Audit Metadata