flyingnobita-agent-profiles
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is primarily a collection of Bash and Python scripts designed to perform system-level operations, including repository management, process scanning, and environment setup.
- Evidence: The main entry point
scripts/agent-profilesdelegates to various sub-scripts inscripts/lib/commands/which perform file operations (mv,rm -rf), process lookups (ps,lsof), and tool execution. - [PRIVILEGE_ESCALATION]: The
setupcommand includes logic to install system prerequisites (e.g.,bubblewrap,git,python3) using system package managers (apt-get,pacman) with elevated privileges. - Evidence:
scripts/lib/mise-tasks.shcontains theap_setup_install_packagefunction, which explicitly prependssudoto commands if the current user is not root:cmd=(sudo ${cmd[@]+"${cmd[@]}"}). - [DATA_EXPOSURE]: The diagnostic (
doctor) and agent-specific modules probe for the existence of sensitive credential files to report on the state of profile isolation. - Evidence:
scripts/lib/agents/agy.shandscripts/lib/agents/claude.shcheck for paths like.claude/.credentials.jsonand.gemini/antigravity-cli/credentials.json. However, the code is carefully written to check for file existence and ownership only, explicitly avoiding reading or printing the actual secret values. - [DYNAMIC_EXECUTION]: The helper script dynamically sources command implementations and agent-specific environment hooks based on the arguments provided.
- Evidence:
scripts/agent-profilesusessource "$cmd_file"to load commands, andscripts/lib/runtime.shiterates through supported agents to load their modules and call environment hooks likeap_agent_env_<id>. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection because it ingests untrusted configuration files from the current Git repository (e.g.,
.mcp.json,settings.json,mise.toml) and has powerful capabilities to mutate the host environment. - Evidence Chain:
- Ingestion points: Reads repository-level configuration files in
scripts/lib/agents/claude.shandscripts/lib/agents/opencode.sh. - Boundary markers: The skill implements its own isolation boundary using
bubblewrap(Linux) and environment redirection (macOS), but these are designed for workflow separation rather than adversarial security. - Capability inventory: Possesses capabilities for recursive file deletion (
rm -rf), file moves (mv), and package installation viasudo. - Sanitization: Uses
python3 -cwithjson.dumpsto safely escape strings for JSON processing and performs regex validation on profile names, though it remains a target for complex instruction poisoning within configuration files.
Audit Metadata