First Seen
Feb 18, 2026
Last Scanned
Feb 22, 2026
Findings
4
Score
85/100
Findings (4)
Detects instructions to hide actions from the user
NEVER share your humanRemove directives that attempt to change the agent's output format, suppress safety warnings, or alter response structure in ways that bypass safety controls.
Likely FP if the output format directive is a legitimate tool configuration (e.g., return results as JSON) that does not suppress safety features.
Detects MCP server configs using npx to run packages without version pinning
"command": "npx"Pin the npx package in the MCP config to an exact version (e.g., @scope/server@1.2.3). Unpinned npx commands can silently fetch a compromised package version.
Likely FP if the MCP config is a local development setup example, though unpinned npx in production configs is a real supply chain risk.
Detects -y, --yes, or --auto-approve flags in MCP/skill install commands that bypass user confirmation
"-y"Remove the -y/--yes auto-confirm flag from MCP server launch arguments. This flag bypasses user confirmation prompts and allows unattended execution of potentially dangerous operations.
Likely FP if the matched text is an isolated flag (-y or --yes) in documentation describing command-line options, not in an actual MCP config.
Detects pip install of arbitrary packages that modify the host environment
pip3 install fuPin all pip packages to exact versions (e.g., pip install package==1.2.3). Use a requirements.txt or pyproject.toml with pinned versions and hash verification.
Likely FP if the match is in documentation showing how to install the skill's own PyPI package.