Omniscio

Safety checks (what protects your computer from your agents)

Safety checks are the guard rails between your agents and your computer. Three are always on and need nothing from you: catastrophic commands are refused, approvals nobody is there to answer are refused, and secrets are masked in text that leaves the app. Everything else sits behind one Lab switch, Safety checks, which is off until you turn it on. They lower the chance of a bad accident; they are not a guarantee, and the last section says what they do not cover.

On this page

What it is

When an agent works for you it can run commands, edit files and fetch web pages. The Agent Permission Level (see agent-permission-level.md) decides when Omniscio stops to ask you. Safety checks sit around that choice and catch the cases a level alone does not: the command that would wipe a disk, the approval card nobody answers, the key that leaks into a message.

There are two groups.

Always on. These work whether or not you ever open the Lab. Each has its own off switch in Settings, Workflow, Agent permissions:

  • Refuse catastrophic commands. Commands that wipe a disk, a home folder or a whole system (for example rm -rf /, a fork bomb, format C:) are refused at every permission level, Full trust included. No approval card is shown, and no approval or allow rule can let one through.
  • Refuse asks nobody is there to answer. If a scheduled job, a recipe, an automation or a hidden session needs approval, the ask is refused at once instead of waiting forever. You get one Needs You alert per command, with an Open session button.
  • Remove secrets before text leaves this computer. Keys, tokens and passwords are masked in text the app sends on an agent's behalf: automation messages, scheduled-job deliveries and errors returned by Omniscio's own tool servers. Commit IDs and share links are left intact so they stay usable.

Behind the Lab switch. Turn on Safety checks in Settings, Lab. A Safety checks section then appears in Settings (it is reached by search or a deep link, not from the side list), and extra cards appear under Workflow, Agent permissions. With the switch off, Omniscio behaves as it did before apart from the three always-on protections.

Where to find it

  • Lab switch: Settings, Lab, Safety checks.
  • Safety checks section: one switch or choice per check, each with help text and a search entry. While the Lab switch is off the section shows an off card.
  • Always-on switches: Settings, Workflow, Agent permissions.
  • Smart level, never-allow rules, suggested rules: Settings, Workflow, Agent permissions (the Lab switch must be on).
  • Approval card: the card above the message box shows Allow for this session in the arrow menu beside Allow, and the keyboard shortcut Shift+S does the same while the card is up.
  • Status for tools and agents: GET /safety/status on the local control server (read only).
  • Suggested rules from a terminal: GET /safety/rule-suggestions, then POST /safety/rule-suggestions/<id>/accept or /dismiss, with your own control-server token.

How it behaves

The Lab-gated checks, one by one

Every row below needs the Lab switch on. Defaults are what you get the moment the switch goes on.

Check What it does Default
Ask before dangerous commands Force-pushes, recursive deletes outside the project, world-writable permissions, piping a download into a shell, database drops, stopping system services, shutdown and similar stop and ask, even at Autonomous and Full trust. Not as bad as the always-refused list, so you can still say yes. On
Block writes to credential folders Agents cannot write into folders that hold keys and logins (SSH keys, cloud credentials, Kubernetes and GPG folders, package-registry and git credential files) at any level, Full trust included. On
Keep secrets out of agent processes Strips environment variables whose names look like keys, tokens, secrets, passwords or credentials before an agent's command or a scheduled script starts. On
Environment variables to let through The names you list stay visible even with the scrub on (up to 50). Empty
Allow agents to reach private network addresses When off, agent web fetches and web-style tool servers are refused for your own computer's network side, your home or office network and cloud-metadata addresses. Off (blocked)
Apply the site blocklist to every agent web tool Sites you blocked or marked no-access for Agent Web are also refused by other fetch and search tools. On
Check project instruction files Scans CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, the Copilot instructions file and the .claude/rules set for text that tries to take over the agent. On
Check memory edits and imports Scans new and edited memories and imported memory files. On
Check scheduled prompts Scans the prompts of scheduled jobs and recipe schedules. On
Warn about known-vulnerable packages Daily lookup of known security problems. Off
Suggest rules from my approvals Offers rules for commands you keep allowing. On
Unanswered approvals Waits for you, or refuses after 5, 15 or 30 minutes. Wait for me
Smart approval daily spending limit Caps what the Smart level may spend in a day. $0.50 (range $0.05 to $10)

The Smart level, never-allow rules and Allow for this session are also part of the Lab group; they have no switch of their own beyond it.

Smart approvals

Smart is a fifth permission level, offered in the level dropdown only while the Lab switch is on. It decides exactly what Guarded decides: anything Guarded would let through still goes through, and everything Guarded would ask about goes to a small background model instead.

  • The model sees the tool name, the command or path with secrets removed (at most 2,000 characters), the project folder's name, any dangerous-command match and your level, all treated as plain data.
  • It answers allow, deny or ask, plus a one-line reason (200 characters at most). Anything else counts as ask. Each answer leaves a line in the session, such as "Smart approval allowed npm test: runs the project's tests".
  • It has 8 seconds. No key, a timeout, a garbled answer or a spent daily limit all mean ask you. Smart never allows on failure.
  • A command that matches the dangerous list can only be denied or sent to you, never allowed by the model.
  • The same command in the same session reuses its answer for 30 minutes, and a remembered deny is never turned into an allow.
  • Cost cap: every call is billed as a "smart-approval" cost to the asking session. When the daily cap (Safety checks section) is reached, Smart sessions ask you until tomorrow.
  • What it never allows: the always-refused list, your never-allow rules and credential-folder writes are decided before the model is consulted, and so are the locks on reviewer, tool, crew, email and Clean Room sessions. The model never sees them. Turning the Lab switch off makes a Smart session behave as Guarded.
  • Smart maps to Guarded-style asking on Codex.

Never-allow rules

A never-allow rule is a pattern an agent is always refused, at every level including Full trust and Smart, with no approval card. You add them in Settings, Workflow, Agent permissions (up to 100 rules, 200 characters each). An agent can ask for a change over the control server, which raises a card only you can approve. Nothing else outside that screen can write them.

  • Command rules. Type the command. A rule ending in * matches anything that starts the same way, so git push --force* refuses every force-push and rm -rf * refuses every recursive delete. Without the * it matches only that exact command.
  • Path rules. Start with path: to refuse a file or folder instead, for example path:~/secrets/**. A path: rule needs a path after it.
  • Which engines get them. The built-in engine (local and cloud sessions) and the external agents that use the shared approval channel (Kimi Code, Hermes, DSH) get every rule. Codex gets a copy as a rules file Omniscio writes into each Codex home it launches (omniscio-safety.rules; Omniscio never touches another rules file). Only command rules written as a prefix ending in * can be written that way. Exact rules, path: rules and patterns with a * in the middle are not applied to Codex. Pi is not reached at all.

Allow for this session

On an approval card, Allow for this session (arrow menu beside Allow, or Shift+S) allows that request and stops asking about the same command until the session ends. It matches the way "always allow" rules do (the command's leading words; chained commands do not slip through). The grant lives in memory only and is gone when the session exits, restarts or the app closes. It never covers an always-refused command, a never-allow match or a credential folder. It is offered on the built-in engine and on Codex (where it uses Codex's own "for the session" answer), and on external agents only where their card can offer it.

Unanswered approvals

  • Nobody there: a session started by a schedule, recipe, workflow, automation rule or trigger, or one started hidden, is "unattended". Its asks are refused immediately with the reason "nobody was there to approve it", and you get one Needs You alert per session and command.
  • You are there but busy: the Unanswered approvals choice sets an optional timeout (off by default). When it runs out the request is refused, the session is told, and a later answer is ignored.

Credential folders and .env files

Writes into home credential folders are refused at every level, including shell commands that write there by redirection, copy, move or a download's output flag (curl -o, wget -O, -OutFile). A project .env or .env.<name> file is different: writing one asks first, at every level below Full trust. Files named .env.example, .env.sample, .env.template and .env.defaults are never flagged. On external agents the existing behaviour (refuse) is kept. On Codex only shell commands the rules file can match are caught.

The secrets scrub and passthrough

With the scrub on, every agent session's child processes and every scheduled script start without environment variables that look like secrets. The following are kept: the agent engine's own provider keys, Omniscio's own session tokens and the password for the mail tool. If a tool you run really needs one, add its exact name to Environment variables to let through. Names must be letters, digits and underscores and may not start with a digit.

Outbound masking

Always on. Masking uses one shared list of secret shapes and applies to automation messages, scheduled-job deliveries and errors from Omniscio's own tool servers. It cannot be bypassed by a Lab switch because it is not behind one.

Private networks and blocked sites

  • Private network. Addresses that point inside your computer or network (private ranges, link-local, carrier-grade NAT, cloud metadata) are refused for agent web fetches and for web-style (HTTP/SSE) tool servers, both when added and when loaded into a session. Your own computer (loopback) stays allowed for tool servers, as before. Turning on Allow agents to reach private network addresses opens the private ranges, but cloud-metadata addresses are always refused.
  • Site blocklist. Agent Web's blocked and no access site levels also apply to the built-in engine's web fetch tool and to the app's own fetches, including after a redirect.
  • Shell curl is not covered. A command an agent runs in a terminal, such as curl or wget, is a shell command. It is neither checked against your blocklist nor stopped by the private-network rule. Only the "download into a shell" pattern is caught, by the dangerous-command ask.

Hidden-instruction scanning

The scanner looks for text that tries to give an agent orders (for example "ignore previous instructions") and for secrets pasted in. It reports reasons by label and never repeats the text.

  • Instruction files. A flagged file is still loaded, but fenced and marked as untrusted data, and you get one Needs You alert per file version with Open file. An oversized file (over 512 KB) is skipped by the check.
  • Memory. Edits to a memory's title or tags are scanned like an agent's own saves. When memory files are imported, a flagged item is held instead of imported; the import dialog offers Import anyway.
  • Scheduled prompts. The prompt is checked when you create the job (it is still saved, with a warning) and before each run. A flagged prompt's run does not start; a Needs You card offers Run anyway, Skip this run and Edit job. Run anyway remembers that exact prompt text, so later runs with the same text go ahead; edit the prompt and it is checked again.

Package advisories (off by default)

Once a day (and shortly after start-up) Omniscio can look up known security problems for the installed engine tools that come from npm and the npx packages in your tool-server setup. It talks to the public database api.osv.dev and sends only package names, versions and the ecosystem (npm): no file paths, project names, keys or account details. A package flagged as malicious, or with a critical problem, raises a Needs You alert naming it and the fixed version. If you are offline the check is skipped and retried the next day. It does nothing until you turn Warn about known-vulnerable packages on, and it is listed in the project's outbound-hosts record as an optional, non-personal destination.

Suggested rules

When you allow the same kind of command by hand at least 3 times in 30 days across at least 2 sessions, a suggestion appears under Workflow, Agent permissions, reduced to the command's leading words. Add as a rule creates the allow rule; Dismiss hides it. Nothing is added on its own.

Decision history

Every approval decision is recorded: who decided (you, a rule, a session grant, Smart, the unattended refusal, a timeout, a built-in refusal or a never-allow rule), the answer, the level, the engine and the command with secrets removed (500 characters at most). Rows older than 90 days are deleted daily. The history feeds suggested rules and the 7-day count in the status route. It is part of data export, is removed with its session or project, is erased by the account's erase-all, and is left out of backups because it can be rebuilt over time.

What it does not protect against

  • Engines Omniscio cannot reach. Pi has no point where Omniscio can intercept a command, so none of this applies there. Codex gets only the command half described above.
  • Shell network tools. curl, wget and similar are not subject to the site blocklist or the private-network rule.
  • A command that looks harmless. The lists match patterns. A determined or lucky script can hide a bad action in a form no pattern recognises, and Full trust approves everything outside the always-refused list.
  • Smart is a judge, not a promise. It can be wrong; its reason lines let you audit it, and failures always come back to you.
  • Secrets in places we do not scan. Masking covers the listed outbound paths. A key you paste into a file an agent then uploads yourself is not caught.
  • Your own approvals. If you click Allow, the command runs as written.
  • Cloud machines and containers beyond the session. This is about your agents' sessions, not hardening of servers.

Last verified 2026-10-28