cotty promises that what a person types at a secret prompt never reaches the agent, that values on the masking list never reach it either, that the person's view of the session is honest, and that an agent at a given assurance level cannot do what docs/security.md says that level stops. A way around any of these is a security problem. Please report it privately and give us time to fix it before it is disclosed.
Use GitHub's private vulnerability reporting for this repository: the "Report a vulnerability" button under the Security tab (https://github.com/openbasalt/cotty/security/advisories/new). If you cannot use GitHub, write to security@openbasalt.org.
Do not open a public issue, pull request or discussion for a security problem.
Examples of what we want to hear about:
- secret input, a listed value or text typed in private mode that reaches an agent connection, the MCP server, a log or the audit file;
- an agent that takes the person's role, approves its own command, forges the lock or the status line, or reads or writes the session's terminal;
- an agent under
cotty agentthat reaches something docs/confinement.md says is out of its view, or plants something that runs unconfined later; - an agent tagged by
cotty-guardthat gets past one of its rules, or a session that reportsstrongwhile the protection is not in place.
Known open weaknesses are listed as rows marked [open: ITEM] in the
escape matrix results and in docs/security.md; a report that one of them
is worse than described is still welcome.
Please include what you can of:
- the cotty commit or version, the distribution and the kernel version,
and the output of
cotty doctorandcotty guard facts; - steps to reproduce, and what an attacker gains (who the attacker is: a
prompt-injected agent under
cotty agent, an unconfined program of the same user, another user); - whether you want to be credited, and how.
We aim to acknowledge a report within 7 days and to agree on a disclosure date with you, normally within 90 days of the report or when a fix is released, whichever comes first. Fixes are published with a GitHub security advisory (and a CVE when it applies).
cotty is pre-release. Only the latest commit on main receives fixes.
Run the escape matrix and the guard only in throwaway VMs
(tests/lab/vm.sh), never on a machine you care about.