Markdown. Every skill is a SKILL.md file with no executable scripts, no hooks, no MCP servers, and no network calls. The only code that runs on your machine is scripts/install.sh, which you invoke deliberately, and the maintenance scripts under scripts/ and .github/scripts/.

That is the point: an empirical study of 31,132 public skills found 26.1% contained at least one vulnerability, and skills shipping executable scripts were 2.12 times more likely to be among them. Reading a skill before installing it is a reasonable habit, and here reading it is the whole audit.

• A skill whose guidance would lead an agent that follows it to do something unsafe: exfiltrate credentials, execute untrusted input, disable a control, or take an irreversible action without confirmation.
• Prompt-injection content embedded in a skill, whether deliberate or carried in from a source.
• A flaw in scripts/install.sh that writes outside the target directory or overwrites something it did not create.
• A flaw in the GitHub Actions workflows that would let repository contents or a token be influenced by an untrusted input.

Trading losses are not a vulnerability. These skills teach method; they make no promise about the profitability of any strategy, and a correct method can still lose money.

Report privately through GitHub's Report a vulnerability button under the repository's Security tab, or by email to pm@ml4trading.io.

Please include the file, what an agent following it would do, and how you established that. A working demonstration is welcome; a step-by-step recipe for attacking a third party is not.

You can expect an acknowledgement within a week. Fixes land on main and are noted in the release that follows. There is no bounty.

The tip of main is the supported version. Tagged releases are snapshots and do not receive backported fixes.