planning-with-files-ar
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill automates the injection of project files (
task_plan.md,findings.md) into the agent's context. While this creates a potential surface for indirect prompt injection, it is mitigated by several strong mechanisms: - Plan Attestation: An integrity check system (
scripts/attest-plan.shandattest-plan.ps1) allows users to lock the current plan with a SHA-256 hash. The skill blocks injection if the plan is modified without re-attestation, preventing the agent from being influenced by unauthorized changes to its instructions. - Context Framing: Data is injected using secure delimiters (
===BEGIN-PWF-DATA===) and content-derived nonces. It includes explicit warnings to the agent to treat the content as data only, not as instructions. - Path Containment: The scripts use canonicalization and verified descriptors (e.g.,
O_NOFOLLOWon POSIX) to ensure all file access is contained within the project root, preventing the model from reading arbitrary system files through symlink attacks. - [COMMAND_EXECUTION]: The skill uses local shell and Python scripts to manage its state via platform hooks. These scripts are self-contained and implement security best practices, such as using isolated Python environments (
python -I) and avoiding the dynamic construction of commands from untrusted inputs. - [DYNAMIC_EXECUTION]: The PowerShell implementation of the skill (
scripts/resolve-plan-dir.ps1andscripts/attest-plan.ps1) usesAdd-Typeto compile C# code for calling Windows native APIs. This is a standard and necessary practice for secure path resolution and directory junction handling on Windows. The source code is statically defined within the scripts and serves the primary purpose of the skill.
Audit Metadata