Gate

Severity

Severity is a property of a finding. Blast radius is a property of the agent. They answer different questions and Gate keeps them separate.

The levels

  • Name
    critical
    Type
    fix today
    Description

    A credential has already leaked, or the agent can execute arbitrary code, move money, or reach a credential store. Consequences are immediate and frequently irreversible.

  • Name
    high
    Type
    fix this week
    Description

    The agent can destroy state, administer systems, read secrets, or reach beyond its workspace. Requires a mistake or an injected instruction to become an incident, which is not a high bar.

  • Name
    medium
    Type
    decide deliberately
    Description

    A meaningful weakening of a boundary: an unpinned package, auto-approved tools, a remote server with no visible authentication. Usually a decision somebody should make on purpose rather than by accident.

  • Name
    low
    Type
    worth knowing
    Description

    Process weaknesses rather than exposures. Configuration nobody reviews, for example.

  • Name
    info
    Type
    context
    Description

    Facts worth surfacing that are not problems.

How a rule picks one

Each rule declares a default severity: the one shown by gate rules and on the rule's documentation page. An individual finding may differ when the surrounding configuration justifies it, and Gate says why in the message:

  • GATE014 is high when the production-looking server also exposes write, delete or execute capabilities, and medium when it is read-only.
  • GATE019 is high when the auto-approval covers execution, and medium otherwise.
  • GATE009 is critical when the filesystem grant is the machine root rather than merely outside the repository.

Rules never adjust severity based on how many findings there already are, or on how much you seem to care. The rating for a given fact is the same in every repository.

Severity is not blast radius

A repository can have zero findings and a MODERATE blast radius: nothing is wrong, and the agent can still do a lot. A repository can have one critical finding and a HIGH blast radius, because the one problem is a leaked token on an otherwise tightly-scoped agent.

Severity tells you what to fix. Blast radius tells you how exposed you are. You need both, and conflating them is how security dashboards become decorative.

Choosing a threshold

gate scan --severity high    # the default: fail on high and critical

high is the right starting point. It catches everything that requires action without failing builds over process notes.

Move to medium once the existing findings are dealt with. Going below high means a critical finding will not fail your build, which is almost never what anybody wants. If the goal is to adopt Gate on a repository with existing findings, use a baseline instead of lowering the bar.

Was this page helpful?