Tiltmeter is young. Security fixes go into the latest release only.
| Version | Supported |
|---|---|
| 0.1.x | Yes |
Please do not open a public issue. Report it privately through GitHub: open the repository's Security tab and choose Report a vulnerability. Include what you found, how to reproduce it and what an attacker could do with it.
You can expect an acknowledgement within a week. We will keep you posted while we work on a fix, agree on a disclosure date with you, and credit you in the release notes unless you prefer otherwise.
Tiltmeter runs with your API credentials, talks to third-party APIs and, optionally, to an LLM and to agents. These are the boundaries it is designed to keep:
| Concern | How Tiltmeter handles it |
|---|---|
| Credentials in the config | Kept as ${VAR} placeholders and expanded only inside the request and notification code. Expanded values are never written to the history, passed to the LLM, put in the LangGraph state or logged, and they are scrubbed from error messages and excerpts. The headers of an endpoint go only to the server its url names: a redirect elsewhere is followed without them, and never with the request body. Notifications do not follow redirects. |
| Personal data in responses | Bodies are reduced to shapes on arrival. Only short categorical strings, such as active, are kept. Excerpts are stored or sent to the LLM only when store_samples or llm.include_samples is turned on. |
| Prompt injection through API responses | Responses only reach the LLM as an opt-in excerpt, wrapped in tags and labelled as untrusted data that must not be followed. Explanations are informational: nothing executes them. |
| Agents abusing the MCP server | Tools accept endpoint ids from the config, never URLs, so an agent cannot use Tiltmeter to reach other hosts. The tools that make requests are annotated as such for the client's approval prompts. |
| Webhook URLs | Treated as secrets. Delivery errors never include them. |
| Forged or tampered provider notices | Notices are applied only when an Ed25519 signature verifies against the key file at https://<provider>/.well-known/api-change-key, which must hold nothing but keys and comments, and every endpoint in the notice must be on the provider's domain. Redirects are followed only on https, on that domain and to that path. Unsigned notices are accepted only from a local file with --allow-unsigned. Each notice is handled once. |
| Prompt and markup injection through notices | Notices are validated strictly: no unknown members, no control characters or bidirectional overrides, short text. Their summary and notes reach the agent marked as untrusted data and reach pull requests inside a plain-text block, so they cannot add links, images or mentions. |
| Markup injection through the fix agent | The agent reads repository files and notices, so its own report is untrusted too. Its summary, its open questions and the names of the files it changed reach pull requests as plain text or code, and its summary and title reach the commit only when they hold nothing GitHub would turn into a mention, a link, an image or an issue reference. |
| Running the agent's edits | Only with fix.test_command, which is off by default. The command runs without a shell in the fix's worktree, with no network (a network namespace of its own on Linux, sandbox-exec on macOS, or not at all), only PATH, the locale and the variables you name, and an empty home folder. Its output reaches the agent as untrusted data, and what it writes to the worktree is discarded before the commit. No git hook runs while Tiltmeter restores, commits and pushes, so one that the tests or the agent left behind is never started with your environment and network. It runs as your user, so it is not a sandbox for your files: use it where you would run the repository's tests anyway, such as CI. |
| Large or slow feeds | Feeds, notices and key files are fetched over https only, uncompressed, with a time limit per download and a size limit on the bytes received, and no credentials are sent with them. |
| Cloning strangers' repositories | tiltmeter hunt clones public repositories only to read them as text: one commit deep, with every git hook off, Git LFS downloads off, no credential prompt and a time limit. Nothing in a clone is run, and symbolic links are not followed. Repository names come from GitHub and are checked before they become a folder or a URL, private repositories your login can see are never cloned or listed, and names and file lines are printed as plain text. The command never forks, pushes or opens anything on GitHub. |
| Other people's words on your pull requests | tiltmeter prs reads titles, logins, check names and comments from pull requests anyone can write on. It strips control characters, cuts them to one line and prints them as plain text, never as markup, and shows only the start of a comment. The command it suggests takes a check or notice id from a pull request's footer only when the id has the shape of one, names a notice only when the bundled catalog has that entry, and is quoted for the shell. Logins and repository names are checked before they become part of a search. The command only reads: it never comments, pushes, merges, closes or reruns anything. |
| The history database | A local SQLite file with no credentials. Protect it like any other local data, and keep .tiltmeter/ out of version control. |
| Supply chain | Releases are built in GitHub Actions and published to PyPI with trusted publishing, without long-lived tokens, together with build provenance attestations. Workflow actions are pinned to commit SHAs. |
Out of scope: the configuration file itself is trusted input. Anyone who can edit it can point Tiltmeter at any URL with your credentials, so review changes to it like code.