DeepSeek Harness Plugins · Security

Security plugins on this shelf are the defensive ones: checks around what the agent can run, scanners for a diff, warnings before a dangerous command. They are a fit when you already have an agent that can act and you want a brake. They are not a fit as a first install on a machine where you have not decided what the agent is allowed to do. Write that allow-list in your own words first. A plugin cannot invent your policy.

Read the verbs. Explain, block, and log are defensive. Flood, evade, and generate payloads are not this category even if a card says security in the title. Those belong under admin-security, and they belong in a lab you are allowed to test. If our chip and the README disagree, believe the README and tell us through the contact form.

A security plugin with shell access can lock you out. Know the bypass and have a console that does not depend on Harness. Run the first test on a scratch repository with a deliberate mistake, such as a committed test key you created for the exercise. The plugin should flag it. If it does not, it is not protecting the real repository either.

Do not confuse this shelf with IT skills in the skill directory. A skill is a prompt. A plugin is code in Harness. Installing a prompt that talks about iptables does not install this plugin.

Stars are a particularly bad proxy here. Attention gathers on dramatic claims. A quiet check that catches one class of mistake is the plugin you can live with. Prefer the quiet one if you can see the check fail and then pass.

Run the check on a repository that contains the planted mistake, then on a clean one. You want a fail and a pass. A plugin that passes both has not shown you anything. Do not point it at production until you have seen it catch the planted key.