agent-host-isolation
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill is a defensive tool designed to improve AI safety and host security by enforcing guest VM isolation, resource limits, and credential brokers.
- [COMMAND_EXECUTION]: The skill includes a local validation script
scripts/validate-manifest.py. This script performs static analysis of JSON manifests to ensure they follow security best practices (e.g., read-only mounts, denied network by default). It does not execute arbitrary code or perform unsafe operations. - [INDIRECT_PROMPT_INJECTION]: While the skill instructions handle untrusted data (Step 1, Item 2), the entire procedure is designed to mitigate risks from such data by moving execution to a restricted guest environment. It does not interpolate untrusted data into sensitive prompts without structural boundaries.
- [DATA_EXFILTRATION]: The skill explicitly instructs users to deny network access by default and route all external side effects through a host-side "result gate" to prevent unauthorized data transfer.
- [CREDENTIALS_UNSAFE]: The skill correctly identifies credential exposure as a risk and mandates the use of task-scoped brokers instead of passing environment variables or SSH agents to the agent's execution environment.
Audit Metadata