nx-multi-repo-migrate
Fail
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill provides explicit instructions to disable the agent's security sandbox ("disable the sandbox + restart") if errors occur during package installation with pnpm. This encourages the user to lower global security boundaries to facilitate tool operation.
- [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill processes instructions from files located at
tools/ai-migrations/**/*.mdand<nx_migrate_prompt>blocks provided in the output of thenx migratecommand. - Boundary markers: There are no defined delimiters or instructions for the agent to treat these external prompts as untrusted data.
- Capability inventory: The agent possesses
Bash,Write, and specializedpolygraphagent tools, which allow it to modify source code and execute shell commands based on the external instructions. - Sanitization: The skill lacks any sanitization or validation mechanisms, instructing the agent to "read each prompt and make the described changes" directly.
- [CREDENTIALS_UNSAFE]: The documentation describes techniques to manipulate sensitive environment variables like
GH_TOKEN(e.g.,env -u GH_TOKEN) and suggests modifications to global configuration files such as~/.npmrc,~/.config/pnpm/rc, and~/.yarnrc.ymlto bypass security-related release-age gates. - [COMMAND_EXECUTION]: The skill implements a workflow where it parses shell commands from external tool output (specifically
<nx_migrate_step>blocks) and directs the agent to "Run the command verbatim," leading to dynamic execution of commands provided by a external source. - [EXTERNAL_DOWNLOADS]: The skill relies on
nx migrateto fetch and execute migration scripts and dependencies from public registries at runtime. While these originate from the vendor's ecosystem, they constitute remote code execution of unverified scripts. - [PROMPT_INJECTION]: The instructions include directives to the agent to mask its identity, such as "never mention AI/Claude" in commit messages, which is a form of instruction override to hide the agentic nature of the operations.
Recommendations
- AI detected serious security threats
Audit Metadata