English | 简体中文

Please report security issues privately, through GitHub's private vulnerability reporting:

Security tab on this repository → Report a vulnerability (direct link: https://github.com/angel291592/Intent-Router/security/advisories/new).

Do not open a public issue for a security problem. You will get an acknowledgement, and the report stays private until a fix is available.

Intent-Router is a prompt-only skill: a markdown file plus reference documents and a JSON Schema. There is no runtime, no server, and no code that handles untrusted input on its own. The realistic issue classes are:

• A prompt-injection path where the skill's instructions cause an agent to take a harmful action (for example, reading or exfiltrating something the user did not intend).
• Instructions that cause an agent to write outside the workspace, or to act on an irreversible boundary without asking. The skill's only write is its own spec file under .intent/ — a saved spec quotes the original request and the record values it probed, so in non-code domains it can carry customer data; whether .intent/ is committed or ignored is the project's own decision.
• A weakness in the evaluation runner (evals/run.py) when pointed at untrusted fixtures.

• The behaviour of a specific harness or model. Report those as regular bugs, or upstream.
• The fixtures under evals/fixtures/ (both of them): deliberately small probe targets, not shipped applications.

Only the latest released version is supported. Fixes land on main; there are no maintained release branches yet.