ai-scaffold-dev-context

Pass

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from various repository files, which creates an attack surface for indirect prompt injection.
  • Ingestion points: The skill reads several project files including README.md, AGENTS.md, CLAUDE.md, CONTRIBUTING.md, and CI configuration files like .gitlab-ci.yml or .github/workflows/.
  • Boundary markers: The SKILL.md file contains a robust "Untrusted Content" section that explicitly instructs the agent to treat any imperatives found in read content (e.g., "ignore previous instructions") as findings to be reported rather than instructions to be followed.
  • Capability inventory: The skill uses the Bash tool to perform git and filesystem queries (git rev-parse, git ls-files, test) and has the capability to write a diff back to the repository's root documentation files.
  • Sanitization: Step 4 of the skill procedure implements a "withheld" logic that redacts internal hostnames and project identifiers if the repository is determined to be public, mitigating accidental data exposure.
  • [PROMPT_INJECTION]: The skill instructions contain phrases often flagged as prompt injection, such as "ignore previous instructions."
  • Evidence: These phrases are used as negative examples within the "Untrusted Content" defensive block to teach the agent what types of imperatives it must report and decline.
  • Assessment: This is a defensive instruction designed to improve the agent's resilience against external attacks, not an attempt to bypass system safety guidelines.
  • [COMMAND_EXECUTION]: The skill executes shell commands via the Bash tool to extract repository metadata.
  • Evidence: It uses commands like git rev-parse --show-toplevel, git remote get-url origin, and git ls-files to gather facts about the forge and repository structure.
  • Mitigation: The skill's frontmatter configuration strictly limits the Bash tool to a whitelist of read-only git subcommands and filesystem tests, while explicitly disallowing dangerous operations like git push.
  • [DATA_EXFILTRATION]: The skill collects and records environment-specific data, including forge hostnames and project identifiers.
  • Evidence: The skill records web_host, ssh_host, project, and project_id into a local AGENTS.md file.
  • Mitigation: The skill includes automated logic to detect the visibility of the project and will redact internal infrastructure details if the repository is destined for a public target, preventing accidental information leaks.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 9, 2026, 02:44 PM