Security

How SKaaS isolates tenants, protects tokens, treats skill content as untrusted data, and what is accepted risk, with the limits stated plainly.

By The SKaaS team. Last updated . Markdown version

This page describes the controls that SKaaS uses and the limits of what they show. It is not a security assessment, and SKaaS makes no compliance claims.

How are organizations kept apart?

Every query carries the organization and the caller's grants, and results are checked again after the database returns them. A known identifier never bypasses a check. A skill that is archived, quarantined or yanked is refused according to its state, whatever the caller knows about it.

How are tokens handled?

Tokens are stored hashed. They carry scopes (skills:read, skills:search, skills:write, repos:admin), can be limited to named repositories, have an expiry and a last-used time, and you can revoke one at any time. A token is shown once, when you create it. A key for a person starts with skaas_pat_ and a key for a service with skaas_svc_. A personal key never reaches more than its person does. Revoking a key, removing a grant, deactivating a member, quarantining a skill and withdrawing a release for security each take effect at once on the server that handles the change, and within 60 seconds on every other server, because cached access decisions expire that fast.

Stored secrets that the service must read back, such as the signing secret of a webhook, the private signing keys of an organization, the key for the Claude Skills API and the model key of an organization, are sealed with encryption keys that the deployment holds. A person can see only the last four characters of such a secret.

How should I treat skill content?

As untrusted data. A skill can contain instructions that try to redirect an assistant. The auditor looks for known patterns and shows its result next to each skill, but it cannot prove a skill safe. Load skills from publishers you trust, pin releases, and read changes before you update. See the audit and its limits.

What does the prompt hook send?

By default, nothing leaves your machine. The remote fallback is off unless you turn it on, and it sends at most 2,000 characters of a prompt.

What is stored about searches?

The server never logs skill content, prompts, tokens or connection strings, and it stores no local file paths. An organization whose plan has search analytics has a search log. By default the log keeps a hash and the length of each query and not its words. It keeps the text only when an administrator switches on the setting search_log_query_text. An administrator can also switch the log off with search_log_enabled, and an entry expires with the audit log retention of the plan. The hash is a keyed hash (HMAC-SHA-256) of the case-folded query, under a key that the deployment derives from its authentication secret, so a person who can read the log cannot test a guess against it without that key. The length is the number of characters of the query. The switch has three positions, off, hash and length only, and words too, on the Search page and in the settings of the dashboard.

What is accepted risk?

The hosted service runs on Cloudflare Workers, which have no fixed outgoing address, so the database network allow list is open and the controls are TLS, least-privilege database users, rotated credentials, auditing and alerts. The least that the service needs from its database user is readWrite on the application database and one custom role, skaasSchema, which allows collMod on that database and nothing else, because the server sets the validator of every collection when it starts. A user with those two roles cannot drop the database, bypass a validator or reach another database. The self-hosted stack creates its application user with exactly these roles, and the runbook on rotating database credentials creates a hosted user with them. Customers who cannot accept this can run a private data plane, which is an Enterprise option.

How do I report a problem?

No security reporting address is published yet. The SECURITY.md file in the repository says what to include in a report and what is in scope, and this page will name the channel when it exists. If a key or a token leaked, the runbook on a leaked token in docs/runbooks/ of the repository lists the steps.

Public guide and resource responses, including pinned reads, may stay in a shared cache for up to one minute after a security withdrawal or takedown. The origin refuses new reads immediately. There is no cache-purge operation. Recheck a pinned read after that minute.