Skip to main content
Commands, package names, and image names on this page come from the open-source project that Mibyan Desktop is built on, and can differ from the Mibyan Desktop installer. For the supported Mibyan install and update path, see Install and update.
Mibyan is designed with a defense-in-depth security model. This page covers every security boundary — from command approval to container isolation to user authorization on messaging platforms.

​
Overview

The security model has eight layers:
  1. User authorization — who can talk to the agent (allowlists, DM pairing)
  2. Dangerous command approval — human-in-the-loop for destructive operations
  3. File write safety — denylist and optional write sandbox for write_file/patch
  4. Container isolation — Docker/Singularity/Modal sandboxing with hardened settings
  5. MCP credential filtering — environment variable isolation for MCP subprocesses
  6. Context file scanning — prompt injection detection in project files
  7. Cross-session isolation — sessions cannot access each other’s data or state; cron job storage paths are hardened against path traversal attacks
  8. Input sanitization — working directory parameters in terminal tool backends are validated against an allowlist to prevent shell injection

​
Dangerous Command Approval

Before executing any command, Mibyan checks it against a curated list of dangerous patterns. If a match is found, the user must explicitly approve it.

​
Approval Modes

The approval system supports three modes, configured via approvals.mode in ~/.mibyan/config.yaml:
The full set of keys:
Setting approvals.mode: off disables all safety prompts. Use only in trusted environments (CI/CD, containers, etc.).

​
YOLO Mode

YOLO mode bypasses all dangerous command approval prompts for the current session. It can be activated three ways:
  1. CLI flag: Start a session with mibyan --yolo or mibyan chat --yolo
  2. Slash command: Type /yolo during a session to toggle it on/off
  3. Environment variable: Set mibyan_YOLO_MODE=1
The /yolo command is a toggle — each use flips the mode on or off:
YOLO mode is available in both CLI and gateway sessions. Internally, it sets the mibyan_YOLO_MODE environment variable which is checked before every command execution. When YOLO is active, Mibyan shows two persistent visual reminders so it’s hard to forget that approval prompts are bypassed:
  • A red banner line at session start when YOLO is already active: ⚠ YOLO mode — all approval prompts bypassed. Hidden when YOLO is off so the default banner stays uncluttered.
  • A ⚠ YOLO fragment in the status bar across all width tiers, updated live as you toggle YOLO on or off (rich-text renderer and plain-text fallback).
YOLO mode disables all dangerous command safety checks for the session — except the hardline blocklist (see below). Use only when you fully trust the commands being generated (e.g., well-tested automation scripts in disposable environments).
For destructive session slash commands (/clear, /new / /reset, /undo, /quit --delete — /exit --delete is an alias), the CLI also prompts for confirmation before running them. See Slash Commands — Confirmation prompts for destructive commands.

​
Supervised-gateway lifecycle restriction

The terminal tool has a separate, non-overridable guard against stopping or restarting the gateway from inside its own supervised process. A self-restart can terminate the tool before it finishes and cause a supervisor/auto-resume loop. User approval, YOLO mode, and force=True do not bypass this guard. The guard also refuses process killers aimed at the interpreter image the gateway runs as — taskkill /F /IM python.exe, taskkill /FI "IMAGENAME eq python.exe", Stop-Process -Name python, pkill -9 python3, killall python, pkill -f python, and name-derived kills such as pgrep python | xargs kill — because a supervised gateway is literally a python process and such a command takes it (and the agent’s own turn) down. Kills scoped to a process the agent owns pass: the proc_* id of a background job (process(action="kill", …)) or an explicit PID (taskkill /F /PID <pid>, kill <pid>). Other image names (taskkill /F /IM notepad.exe) are unaffected. The guard is active under every generated launcher — systemd unit, launchd plist, s6 run script and the Windows Scheduled Task — via the mibyan_SUPERVISED_CHILD marker they export. On macOS, executed launchctl submit and launchctl bootstrap commands are restricted regardless of the job label. This is a conservative registration restriction intended to catch indirect restart helpers with neutral labels, not an inspection of the target plist. It also rejects independent scheduled jobs with RunAtLoad=false and no KeepAlive key; rejection does not establish that the job uses KeepAlive or controls Mibyan. For authorized LaunchAgent maintenance, use a separate shell outside the running gateway. Some independent load/unload commands currently pass the label-based checks, but that is not a target-verified exemption or a supported way to evade a bootstrap rejection. Read-only launchctl print is not a lifecycle operation. After external maintenance, distinguish the on-disk plist from the loaded job: validate the plist and read back the loaded schedule before reporting activation. A tool rejection means the command did not execute through that tool call. An assistant declining to issue a call is a separate model decision; changing models does not change the terminal guard’s policy.

​
Hardline Blocklist (Always-On Floor)

Some commands are so catastrophic — irreversible filesystem wipes, fork bombs, direct block-device writes — that Mibyan refuses to run them regardless of:
  • --yolo / /yolo toggled on
  • approvals.mode: off
  • Cron jobs running in headless approve mode
  • User explicitly clicking “allow always”
The blocklist is the floor below --yolo. It trips before the approval layer even sees the command, and there’s no override flag. Patterns currently covered (not exhaustive; kept in sync with tools/approval.py::UNRECOVERABLE_BLOCKLIST): If you hit the blocklist, the tool call returns an explanatory error to the agent and nothing runs. If a legitimate workflow needs one of these commands (you’re the operator of a wipe-and-reinstall pipeline, for example), run it outside the agent. The floor also fails closed on a command whose shell quoting cannot be parsed (grep 'unterminated): the error says malformed executable payload. Quoting is judged on the command exactly as written, so shell-valid escapes inside a quoted pattern (grep -o "[^\"]*" file) are not malformed, and an escaped quote before a separator (echo "a\"b"; reboot) does not hide the command that follows it.

​
User-Defined Deny Rules (approvals.deny)

The hardline blocklist is fixed and code-shipped. approvals.deny is its user-editable counterpart: a list of glob patterns that block matching terminal commands unconditionally — before --yolo, /yolo, and approvals.mode: off are consulted. Use it to run yolo-with-exceptions: “let the agent do everything, except these specific things, ever.”
Details:
  • Patterns are fnmatch globs (*, ?, [...]) matched case-insensitively against the whole command text and individual executable-command candidates. git push --force* matches git push --force origin main but not git push origin main.
  • Matching runs over the same normalized/deobfuscated command variants the dangerous-pattern detector uses, so simple quoting tricks (git pu""sh --force) don’t slip past a rule.
  • Executable candidates retain the literal path and also match its basename: sudo * covers /usr/bin/sudo -n id and ./sudo -n id. A path-specific rule such as /usr/bin/sudo * does not become a rule for every binary named sudo.
  • Quote-aware parsing exposes commands after assignments, leading redirections, ;, &&, ||, pipelines, groups, command substitutions, and ordinary if/then/else/do transitions. Supported launchers include sudo, env, command, exec, nohup, setsid, time, nice, timeout, stdbuf, ionice, chrt, taskset, and chroot. Known option operands are skipped; command -v/-V lookups are not executions. Shell -c payloads are inspected recursively. Literal executable-and-argument strings in env -S / --split-string use GNU quoting and escapes (including \_ word boundaries and \c termination), with the remaining command arguments appended; shell punctuation inside those arguments stays data unless an actual shell -c consumes it. env -a / --argv0 values are arguments, not executable names. Shell and GNU split-string comments do not introduce executable candidates.
  • In the additional executable candidates, whitespace between words is collapsed, but quoted argument content and argument paths are retained. An exact rule such as git status therefore also matches env git\tstatus; echo done (where \t represents a tab). Quoted mentions such as echo 'sudo -n id' are not promoted to commands. Existing whole-input globs such as *sudo* still intentionally match mentions anywhere.
  • YAML quoting: always quote patterns. A bare leading * is a YAML alias and fails to parse; {, !, and : have their own YAML meanings. Single quotes are safest for shell-ish content.
  • User-defined deny rules apply to all terminal backends, including isolated containers, before any backend-specific approval shortcut.
  • A denied command returns a BLOCKED error to the agent telling it not to retry or rephrase. Nothing runs.
Like the rest of the approval config, changes take effect immediately (the config cache is mtime-keyed) — no session restart needed.
Threat modelDeny rules are a shell-command policy, not a complete shell interpreter or an OS capability sandbox. Normalization does not resolve arbitrary variables (including GNU env -S ${NAME} expansion), aliases, functions, renamed binaries, scripts, interpreter programs, or every shell/launcher grammar (for example, case-pattern syntax, clustered launcher options, or options embedded inside an env -S string). Do not use a basename deny rule as a guarantee that a capability cannot be reached by other means. For containment, use OS permissions and an isolated backend with appropriately restricted mounts, credentials, and network access. This matching behavior does not change the configured approval mode or the empty-deny-list default.

​
Approval Timeout

When a dangerous command prompt appears, the user has a configurable amount of time to respond. If no response is given within the timeout, the command is denied by default (fail-closed). An expired prompt cannot be reopened: the pending entry is discarded and the agent is told not to retry on its own within that turn. To run the operation after all, send a new message asking for it (for example “go ahead and run that now”) — the agent issues a fresh tool call, which raises a fresh approval card, and a “once” approval applies only to that call. A timeout is not counted as a denial, so asking again is never penalized. Configure the timeout in ~/.mibyan/config.yaml:

​
What Triggers Approval

The following patterns trigger approval prompts (defined in tools/approval.py):
Container bypass: When running in docker, singularity, modal, daytona, or vercel_sandbox backends, dangerous command checks are skipped because the container itself is the security boundary. Destructive commands inside a container can’t harm the host.

​
Approval Flow (CLI)

In the interactive CLI, dangerous commands show an inline approval prompt:
The four options:
  • once — allow this single execution
  • session — allow this pattern for the rest of the session
  • always — add to permanent allowlist (saved to config.yaml)
  • deny (default) — block the command

​
Approval Flow (Gateway/Messaging)

On messaging platforms, the agent sends the dangerous command details to the chat and waits for the user to reply:
  • Reply yes, y, approve, ok, or go to approve
  • Reply no, n, deny, or cancel to deny
The mibyan_EXEC_ASK=1 environment variable is automatically set when running the gateway.

​
Permanent Allowlist

Commands approved with “always” are saved to ~/.mibyan/config.yaml:
These patterns are loaded at startup and silently approved in all future sessions. Entries can be exact command text, a shell-style glob (podman *), or a dangerous-pattern rule key such as script execution via heredoc (the key shown in the approval prompt). Rule keys are honored on every surface, including unattended ones: a cron job, mibyan chat -q run or webhook session under cron_mode/single_query_mode/unattended_mode: deny still runs a command whose detected rule key is in command_allowlist, while Tirith content-security findings on the same command continue to block it. The setting must be a list of strings. Legacy installs that stored a list as a quoted YAML/JSON string recover that list at load time and log a warning to re-save it with mibyan config edit. Other malformed values are ignored with a warning; they never become per-character approvals. Loading does not rewrite your configuration file.
Use mibyan config edit to review or remove patterns from your permanent allowlist.
The list is read when Mibyan starts. A pattern you remove while a session is already running stays approved in that session until it next writes the file (the next time you answer always to a prompt) or you restart Mibyan. If you removed it for safety reasons, restart.

​
Mining Approval History (mibyan approvals suggest)

Instead of answering the same prompt session after session, you can mine your past approval decisions into allowlist proposals:
The command scans the session database (~/.mibyan/state.db) for dangerous-classified commands that actually executed — i.e. commands you approved — aggregates them into patterns (git push *, or the dangerous-class key for compound commands), and ranks them by approval frequency:
Safety rules:
  • Nothing is ever applied automatically — the default run is read-only; only an explicit --apply N[,M...] writes to config.yaml.
  • Destructive classes are never proposed, no matter how often they were approved: recursive deletes, sudo, disk/device writes, credential and system-config edits, pipe-to-shell, SQL DROP/TRUNCATE, process kills, and every hardline class are excluded outright. rm -rf build/ approved 100 times still never yields an rm entry.
  • Proposals already covered by your existing command_allowlist are skipped.
  • Credentials inside mined commands are masked (ghp_…, bot<id>:<token> URLs, KEY=value assignments, bearer tokens) in both the printed e.g. examples and the --json payload, using the same redactor as terminal output. Masked text is never used as an allowlist pattern: a command whose glob would embed a credential (TOKEN=… git …) is proposed under its dangerous-class key instead. The session database itself still holds the command as it was executed.
Useful flags: --days N (history window, default 90), --min-count N (minimum approvals to qualify, default 2), --limit N, and --db PATH.

​
File Write Safety

Before write_file or patch touches disk, Mibyan checks the target path against a denylist and an optional sandbox. Blocked writes return an error to the agent immediately — there is no approval prompt and no way to override from the chat UI. The model may still claim the edit succeeded; when display.file_mutation_verifier is on (default), trust the file-mutation verifier footer over the assistant’s closing summary.

​
Protected paths (always blocked)

These categories are always denied, even when mibyan_WRITE_SAFE_ROOT is unset: Project-local .env, .env.local, .env.production and .envrc files are read-denied anywhere on disk (the file tools refuse to read them) but remain writable: the agent can create or edit them for you, it just cannot read the values back. Sensitive paths inside the safe root are still blocked — pointing mibyan_WRITE_SAFE_ROOT at $HOME does not allow writing ~/.ssh/id_rsa. The ~ in the OS-credential rows means every home a write can land in, not just the process HOME: the OS user’s real home, the profile home ({mibyan_HOME}/home under TERMINAL_HOME_MODE=profile, containers and spawned workers, where the process HOME is pinned), and named accounts (~root/.ssh/authorized_keys). An absolute path to the real home’s ~/.aws/credentials is denied even when the agent process runs with HOME pointed elsewhere. Safe-root violations return Write denied: '…' is outside mibyan_WRITE_SAFE_ROOT (…). Credential-path blocks use Write denied: '…' is a protected system/credential file. Exception — ~/.ssh/config is approval-gated, not hard-blocked. The SSH client config holds no private-key material and editing it (host aliases, ProxyJump, VS Code Remote-SSH targets) is a routine task, so write_file / patch route it through the same approve-once/session/always prompt the terminal tool already uses for ~/.ssh writes — instead of the flat refusal that used to apply. It can still carry ProxyCommand / Match exec directives that run commands, so the write is never silent. Non-interactive callers (ACP file bridge, background jobs with no human channel) fail closed. Private keys, authorized_keys, and everything else under ~/.ssh/ remain hard-blocked.

​
mibyan_WRITE_SAFE_ROOT (optional sandbox)

When set, write_file and patch may only target paths inside the listed directory prefix(es). Anything outside is hard-blocked — not routed through dangerous-command approval.
  • Set automatically in the official Docker image (mibyan_WRITE_SAFE_ROOT=/opt/data)
  • Supports multiple roots separated by : on Unix or ; on Windows
  • Do not add to ~/.mibyan/.env casually. If you set it to a project directory, the agent cannot write to ~/.mibyan/cron/jobs.json, profile skills, or other Mibyan state outside that prefix
To allow both a workspace and Mibyan home:
Unset the variable to restore unrestricted writes (subject to the protected-path denylist). Full reference: mibyan_WRITE_SAFE_ROOT.

​
Cron and other Mibyan state

Do not ask the agent to patch ~/.mibyan/cron/jobs.json directly. Use the cronjob_manage tool, mibyan cron, or /cron — they update the job store through the supported API. The same applies to other Mibyan control files when write safety blocks direct edits.
Defense-in-depth, not a hard boundaryWrite guards apply to write_file and patch only, with one exception: the Windows NT/device-namespace row is also enforced on reads — read_file, search_files, @file:/@folder: context references and the ACP file bridge all refuse those paths on the raw string, before anything resolves them. The terminal tool runs as the same OS user and can still cat or overwrite denied paths via shell commands. The denylist reduces accidental damage and gives models a clear stop signal; it does not sandbox a hostile or compromised agent.

​
User Authorization (Gateway)

When running the messaging gateway, Mibyan controls who can interact with the bot through a layered authorization system.

​
Authorization Check Order

The _is_user_authorized() method checks in this order:
  1. Per-platform allow-all flag (e.g., DISCORD_ALLOW_ALL_USERS=true)
  2. DM pairing approved list (users approved via pairing codes)
  3. Platform-specific allowlists (e.g., TELEGRAM_ALLOWED_USERS=12345,67890)
  4. Global allowlist (GATEWAY_ALLOWED_USERS=12345,67890)
  5. Global allow-all (GATEWAY_ALLOW_ALL_USERS=true)
  6. Default: deny

​
Platform Allowlists

Set allowed user IDs as comma-separated values in ~/.mibyan/.env:
The global allow-all can also live in config.yaml as gateway.allow_all_users: true (or top-level allow_all_users: true); a true value is bridged to GATEWAY_ALLOW_ALL_USERS at gateway startup (re-derived on every config load and restart, so flipping it back to false closes the gate), an explicit env var wins, and the gateway logs a warning naming config.yaml as the grant source. In a multi-profile gateway a secondary profile sets GATEWAY_ALLOW_ALL_USERS in its own .env (its config.yaml is never bridged into the process environment).
If no allowlists are configured and GATEWAY_ALLOW_ALL_USERS is not set, all users are denied. The gateway logs a warning at startup:

​
DM Pairing System

For more flexible authorization, Mibyan includes a code-based pairing system. Instead of requiring user IDs upfront, unknown users receive a one-time pairing code that the bot owner approves via the CLI. How it works:
  1. An unknown user sends a DM to the bot
  2. The bot replies with an 8-character pairing code
  3. The bot owner runs mibyan pairing approve <platform> <code> on the CLI
  4. The user is permanently approved for that platform
Control how unauthorized direct messages are handled in ~/.mibyan/config.yaml:
  • pair is the default for chat-style DM platforms. Unauthorized DMs get a pairing code reply.
  • ignore silently drops unauthorized DMs.
  • decline sends one short, polite decline (“I can only chat with my owner”) instead of a pairing code, then ignores further messages from that sender for 24 hours. Customize the text with unauthorized_dm_decline_message.
  • Email defaults to ignore unless platforms.email.unauthorized_dm_behavior: pair is set, because inboxes can contain unrelated unread mail.
  • Platform sections override the global default, so you can keep pairing on Telegram while keeping WhatsApp silent.
Security features (based on OWASP + NIST SP 800-63-4 guidance): Pairing CLI commands:
Docker users: run pairing commands as the mibyan userThe official Docker image runs the gateway as the unprivileged mibyan user (uid 10000) via gosu, but docker exec defaults to root. Approval files created by root are written with mode 0600 root:root and the gateway cannot read them — the approval is silently ignored (#10270).Always pass -u mibyan:
If you already ran the command as root and the user is still unauthorized, restart the container — the entrypoint will fix ownership on the next start.
Storage: Pairing data is stored in ~/.mibyan/pairing/ with per-platform JSON files:
  • {platform}-pending.json — pending pairing requests
  • {platform}-approved.json — approved users
  • _rate_limits.json — rate limit and lockout tracking

​
Container Isolation

When using the docker terminal backend, Mibyan applies strict security hardening to every container.

​
Docker Security Flags

Every container runs with these flags (defined in tools/environments/docker.py):
SETUID/SETGID are not in the base list — they’re added conditionally when the container starts as root and an init/entrypoint must drop privileges (the s6 privilege-drop path). They’re skipped when the container already runs as a non-root --user. The /run tmpfs is also split out from the base list and mounted per-image (hardened noexec by default, exec only for s6-overlay images that exec from /run).

​
Resource Limits

Container resources are configurable in ~/.mibyan/config.yaml:

​
Filesystem Persistence

  • Persistent mode (container_persistent: true): Bind-mounts /workspace and /root from ~/.mibyan/sandboxes/docker/<task_id>/
  • Ephemeral mode (container_persistent: false): Uses tmpfs for workspace — everything is lost on cleanup
For production gateway deployments, use docker, modal, daytona, or vercel_sandbox backend to isolate agent commands from your host system. This eliminates the need for dangerous command approval entirely.
If you add names to terminal.docker_forward_env, those variables are intentionally injected into the container for terminal commands. This is useful for task-specific credentials like GITHUB_TOKEN, but it also means code running in the container can read and exfiltrate them.

​
Terminal Backend Security Comparison

​
Environment Variable Passthrough

Both execute_code and terminal strip sensitive environment variables from child processes to prevent credential exfiltration by LLM-generated code. However, skills that declare required_environment_variables legitimately need access to those vars. First-party platform credentials — the BUZZ_* variables used by the Buzz messaging platform — are passed through to terminal children (foreground and background/PTY spawns) only when the session is actually operating as a Buzz agent: the process is a Buzz-ACP managed agent (BUZZ_MANAGED_AGENT set by the Buzz Desktop harness) or the live gateway session’s platform is buzz. This lets a Buzz platform agent invoke its platform-mandated CLI (e.g. buzz) from the terminal tool, while Telegram/CLI/cron sessions on the same host keep the variables stripped. Because _sanitize_subprocess_env also feeds search workers (e.g. the ddgs web-search subprocess), the computer-use driver binary, and user-script runners (bang ! commands, quick commands, cron scripts, webhook-filter scripts), those children receive the variables too when spawned from a Buzz session. The carve-out is terminal-only: it does not apply to execute_code, browser/TUI-host spawns (mibyan_subprocess_env), Docker/Modal children, or env_passthrough registration, which remain sealed.

​
How It Works

Two mechanisms allow specific variables through the sandbox filters: 1. Skill-scoped passthrough (automatic) When a skill is loaded (via skill_view or the /skill command) and declares required_environment_variables, any of those vars that are actually set in the environment are automatically registered as passthrough. Missing vars (still in setup-needed state) are not registered.
After loading this skill, TENOR_API_KEY passes through to execute_code, terminal (local), and remote backends (Docker, Modal) — no manual configuration needed.
Docker & ModalPrior to v0.5.1, Docker’s forward_env was a separate system from the skill passthrough. They are now merged — skill-declared env vars are automatically forwarded into Docker containers and Modal sandboxes without needing to add them to docker_forward_env manually.
2. Config-based passthrough (manual) For env vars not declared by any skill, add them to terminal.env_passthrough in config.yaml:
Both lists apply to terminal, execute_code and no_agent cron scripts alike. A declared variable is forwarded with the value of the profile the child runs for: when one process serves several profiles (multi-profile gateway, Desktop/dashboard backend) each profile’s declared value comes from its own .env / secret sources, never from the process environment the launch profile populated, and the launch profile’s .env credentials are dropped from a served profile’s children.

​
Credential File Passthrough (OAuth tokens, etc.)

Some skills need files (not just env vars) in the sandbox — for example, Google Workspace stores OAuth tokens as google_token.json under the active profile’s mibyan_HOME. Skills declare these in frontmatter:
When loaded, Mibyan checks if these files exist in the active profile’s mibyan_HOME and registers them for mounting:
  • Docker: Read-only bind mounts (-v host:container:ro)
  • Modal: Mounted at sandbox creation + synced before each command (handles mid-session OAuth setup)
  • Local: No action needed (files already accessible)
You can also list credential files manually in config.yaml:
Paths are relative to ~/.mibyan/. Files are mounted to /root/.mibyan/ inside the container. This list is read by tools/credential_files.py (terminal.credential_files) — it lives under the terminal: block but is loaded by the credential-files module, not the core terminal backend, so it isn’t part of the bundled DEFAULT_CONFIG snapshot.

​
Borrowed CLI logins (Codex CLI, Claude Code)

When Mibyan has no usable login of its own for openai-codex or anthropic, it can borrow the Codex CLI’s ~/.codex/auth.json and Claude Code’s ~/.claude/.credentials.json (or Keychain entry) and refresh them on your behalf. Both use single-use, rotating refresh tokens: once two programs hold one token family, whichever refreshes first invalidates the other’s copy, which shows up as “I logged in once in the terminal and Mibyan keeps failing” (or the reverse). If you run those CLIs alongside Mibyan, give Mibyan its own login and turn adoption off:
With the switch off Mibyan never reads or refreshes those files: the claude_code credential-pool row disappears, mibyan auth list prints one line saying so, and the log carries one INFO line per process. Only automatic adoption is affected — mibyan auth add openai-codex still asks before importing an existing Codex CLI login. Automatic recovery also only repairs the credential Mibyan already holds: a Codex CLI/Desktop login into a different ChatGPT workspace is refused with a warning (re-authenticate with mibyan auth add openai-codex), and a login you complete while recovery is running is never overwritten. Add your own logins with mibyan auth add anthropic / mibyan auth add openai-codex.

​
What Each Sandbox Filters

​
Security Considerations

  • The passthrough only affects vars you or your skills explicitly declare — the default security posture is unchanged for arbitrary LLM-generated code
  • Credential files are mounted read-only into Docker containers
  • Skills Guard scans skill content for suspicious env access patterns before installation
  • Missing/unset vars are never registered (you can’t leak what doesn’t exist)
  • Mibyan infrastructure secrets (provider API keys, gateway tokens) should never be added to env_passthrough — they have dedicated mechanisms. Such a name is refused when declared, and a declared name that a platform adapter claims later (a plugin adapter registering after the skill loaded) stops being forwarded from then on

​
MCP Credential Handling

MCP (Model Context Protocol) server subprocesses receive a filtered environment to prevent accidental credential leakage.

​
Safe Environment Variables

Only these variables are passed through from the host to MCP stdio subprocesses:
Plus any XDG_* variables. All other environment variables (API keys, tokens, secrets) are stripped. Variables explicitly defined in the MCP server’s env config are passed through:

​
Credential Redaction

Error messages from MCP tools are sanitized before being returned to the LLM. The following patterns are replaced with [REDACTED]:
  • GitHub PATs (ghp_...)
  • OpenAI-style keys (sk-...)
  • Bearer tokens
  • token=, key=, API_KEY=, password=, secret= parameters

​
Website Access Policy

You can restrict which websites the agent can access through its web and browser tools. This is useful for preventing the agent from accessing internal services, admin panels, or other sensitive URLs.
When a blocked URL is requested, the tool returns an error explaining the domain is blocked by policy. The blocklist is enforced across web_search, web_extract, browser_navigate, and all URL-capable tools. See Website Blocklist in the configuration guide for full details.

​
SSRF Protection

All URL-capable tools (web search, web extract, vision, browser) validate URLs before fetching them to prevent Server-Side Request Forgery (SSRF) attacks. Blocked addresses include:
  • Private networks (RFC 1918): 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16
  • Loopback: 127.0.0.0/8, ::1
  • Link-local: 169.254.0.0/16 (includes cloud metadata at 169.254.169.254)
  • CGNAT / shared address space (RFC 6598): 100.64.0.0/10 (Tailscale, WireGuard VPNs)
  • Cloud metadata hostnames: metadata.google.internal, metadata.goog
  • Reserved, multicast, and unspecified addresses
SSRF protection is always active for internet-facing use and DNS failures are treated as blocked (fail-closed). Redirect chains are re-validated at each hop to prevent redirect-based bypasses. The same guard covers fetches whose URL comes from a remote party rather than from you: image/video URLs returned by a generation provider, reference-image URLs a model supplies for edits, pet spritesheets and the petdex manifest, and skills.sh sitemap entries. A provider or index that points one of those at a private or metadata address is refused before any connection opens; the operator’s own provider base_url is not affected — a download fetched directly from your configured base_url (the OpenRouter video content endpoint) skips only the private-address class check on that first hop, while the cloud-metadata floor still applies and any redirect it issues is re-validated in full — and an image-generation provider hosted on your LAN needs security.allow_private_urls: true (below) for the result URLs it returns to be cached locally.

​
Intentionally allowing private URLs

Some setups legitimately need private/internal URL access — home networks that resolve home.arpa to RFC 1918 space, LAN-only Ollama/llama.cpp endpoints, internal wikis, cloud metadata debugging, and the like. For those cases there’s a global opt-out:
When on, web tools, the browser, vision URL fetches, and gateway media downloads no longer reject RFC 1918 / loopback / link-local / CGNAT / cloud-metadata destinations. This is a deliberate trust boundary — only enable it on machines where the agent running arbitrary prompt-injected URLs against the local network is an acceptable risk. Public-facing gateways should leave it off. The host-substring guard (which blocks lookalike Unicode domain tricks even when the underlying IP is public) stays on regardless of this setting.

​
Local proxy fake-ip ranges

A TUN proxy in fake-ip mode (Mihomo/Clash fake-ip, Surge enhanced mode) answers DNS with an address from its own block — 198.18.0.0/15 (RFC 2544 benchmarking) by default — for every name outside its filter. Those answers are the proxy’s sentinel, not an internal host, so the private-IP guard otherwise rejects every outbound fetch on such a host: web_extract, platform attachment downloads and the browser relay all fail with URL targets a private or internal network address while the request never reaches the network. Declare the block to let the sentinel through:
Empty by default, and narrower than allow_private_urls: only the declared blocks get the exemption, they should be ranges the local proxy owns (the dial still goes to the proxy, which resolves the real target itself), and loopback, RFC 1918, link-local, CGNAT and cloud-metadata destinations stay blocked — an entry that overlaps one of those classes (including 0.0.0.0/0 or ::/0) is ignored with a warning rather than widening the guard. On a host with a cloud browser provider, the declared sentinel also stops counting as private for browser.auto_local_for_private_urls, so those pages keep going to the cloud browser.

​
Tirith Pre-Exec Security Scanning

Mibyan integrates tirith for content-level command scanning before execution. Tirith detects threats that pattern matching alone misses:
  • Homograph URL spoofing (internationalized domain attacks)
  • Pipe-to-interpreter patterns (curl | bash, wget | sh)
  • Terminal injection attacks
Tirith requests a pinned PM package when enabled and absent. PM checks artifact hashes from pm/lock.json and calls the cosign checker when available. An explicit provenance rejection aborts installation. Startup requests installation in the background, subject to the lazy-install policy. PM owns durable installation state and recovery, not a separate .tirith-install-failed marker. An explicit security.tirith_path remains authoritative, even if the executable is missing. With the default name, lookup uses PATH before the PM selection. External binaries remain outside PM’s hash and provenance checks.
When tirith_fail_open is true (default), commands proceed if tirith is not installed or times out. Set to false in high-security environments to block commands when tirith is unavailable. Three consecutive operational failures (spawn error, timeout, crash) suspend scanning for five minutes so a broken binary cannot stall every command; after that window one command re-probes tirith, and any completed scan (allow, warn or block) resumes normal scanning. A probe that fails again re-arms the five-minute window. PM supports Tirith on Linux (x86_64 / aarch64) and macOS (x86_64 / arm64). With the default path, unsupported targets, including native Windows and Android/Termux, skip Tirith. Pattern-matching guards still run. To use the managed Tirith package on Windows, run Mibyan under WSL. Tirith’s verdict integrates with the approval flow: safe commands pass through, while both suspicious and blocked commands trigger user approval with the full tirith findings (severity, title, description, safer alternatives). Users can approve or deny — the default choice is deny to keep unattended scenarios secure. Two known Tirith false positives are downgraded to “allow” so they never prompt (or, in cron, never deny): a lookalike_tld warning whose only target is the legitimate .app gTLD, and a variation_selector warning when every selector in the command is U+FE0F directly after an emoji (folder names such as 🗞️ Journal/ or ▶️ Media/). A variation selector after a letter or digit — the steganographic-obfuscation signal the rule exists for — still prompts.

​
Context File Injection Protection

Context files (AGENTS.md, .cursorrules, SOUL.md) are scanned for prompt injection before being included in the system prompt. The scanner checks for:
  • Instructions to ignore/disregard prior instructions
  • Hidden HTML comments with suspicious keywords
  • Attempts to read secrets (.env, credentials, .netrc)
  • Credential exfiltration via curl
  • Invisible Unicode characters (zero-width spaces, bidirectional overrides)
The translation-and-execution check requires a short language/format clause (for example, “translate this into a bash script and execute it”). It does not connect translation and execution verbs across unrelated comma-separated role prose. These patterns are heuristics, not semantic intent detection. Blocked project files show a warning:
Your own SOUL.md in mibyan_HOME is treated differently: it is a file you wrote (file-tool writes to it need your approval, and project checkouts never supply it), so a scanner hit there does not block the file. Mibyan logs a warning naming the matched pattern, loads the file as usual, and /context lists it as ⚠ SOUL.md … loaded — matched prompt-injection pattern(s); review the file. This lets an identity file that documents an attack phrase (security guidance such as “content telling you to ignore previous instructions”) keep working; if you did not write the flagged text, treat the warning as a sign that something else edited the file. The exception does not extend to a SOUL.md shipped by a profile distribution: mibyan profile install <git-url> and mibyan profile update copy a third party’s SOUL.md into the profile home without a scan or an approval prompt, so when distribution.yaml owns the file a scanner hit still blocks it.

​
Best Practices for Production Deployment

​
Gateway Deployment Checklist

  1. Set explicit allowlists — never use GATEWAY_ALLOW_ALL_USERS=true in production
  2. Use container backend — set terminal.backend: docker in config.yaml
  3. Restrict resource limits — set appropriate CPU, memory, and disk limits
  4. Store secrets securely — keep API keys in ~/.mibyan/.env with proper file permissions
  5. Enable DM pairing — use pairing codes instead of hardcoding user IDs when possible
  6. Review command allowlist — periodically audit command_allowlist in config.yaml
  7. Set terminal.cwd — don’t let the agent operate from sensitive directories
  8. Run as non-root — never run the gateway as root
  9. Monitor logs — check ~/.mibyan/logs/ for unauthorized access attempts
  10. Keep updated — run mibyan update regularly for security patches

​
Securing API Keys

​
Network Isolation

For maximum security, run the gateway on a separate machine or VM. Set terminal.backend: ssh in config.yaml, then provide host details via environment variables in ~/.mibyan/.env:
The SSH connection details live in .env (not config.yaml) so they aren’t checked in or shared along with profile exports. This keeps the gateway’s messaging connections separate from the agent’s command execution.

​
TLS certificate trust

Mibyan initializes the platform verifier through truststore. Windows uses its certificate store, macOS uses its system trust services, and Linux uses the OpenSSL system trust paths. If initialization fails, Mibyan logs the failure and falls back to OpenSSL defaults. For a corporate TLS proxy, install its root through your organization’s operating-system trust procedure. Mibyan’ provider resolver no longer selects trust through mibyan_CA_BUNDLE or the old CA-environment-variable ladder. Sandboxed subprocesses can have their own separate CA configuration. The former startup certificate guard is gone with it: Mibyan no longer validates mibyan_CA_BUNDLE / SSL_CERT_FILE / REQUESTS_CA_BUNDLE / CURL_CA_BUNDLE at launch, so there is no SSLConfigurationError and the mibyan_SKIP_SSL_GUARD escape hatch has no effect. mibyan_CA_BUNDLE is still honoured by the Nous Portal login flow only (mibyan login, or its --ca-bundle flag); the standard SSL_CERT_FILE / REQUESTS_CA_BUNDLE / CURL_CA_BUNDLE variables are still read by the plain requests/urllib calls some tools make (and by pip, uv, curl, Node), so a stale path in one of them now fails at the call that uses it rather than at startup. Fix or unset the variable there. A custom provider can declare ssl_ca_cert for its endpoint. That bundle replaces platform trust for chat, model metadata, and model catalog probes. A missing file produces a warning and falls back to platform trust. ssl_verify: false disables certificate verification and is unsafe for untrusted networks. Do not use it as a permanent fix for a missing corporate root. Provider HTTP clients keep proxy configuration separate from certificate selection. A stale ambient CA-file path cannot prevent those clients from starting. PM index credentials are sent only to their exact HTTPS origin; redirects to another origin do not receive them.

​
Trusted-by-placement extension points

Most third-party code Mibyan can run is gated by an explicit allow-list: general plugins need plugins.enabled, shell hooks need a first-use approval (or hooks_auto_accept), MCP servers are listed in config. One surface is deliberately different: The gateway imports every valid hook directory in-process, with the gateway’s own privileges. This is the documented contract (since 3988c3c245f), not an oversight: the profile home is operator-owned configuration, and anyone who can write into it can already run code as you through config.yaml shell hooks or by editing plugins.enabled, so a separate consent gate for hooks/ would add friction without moving the trust boundary. Treat the contents of ~/.mibyan/hooks/ like the contents of config.yaml — review a handler.py before you place it, and include ls ~/.mibyan/hooks/ whenever you audit the rest of the profile home (the directory is not on the protected-paths denylist, so it is ordinary writable state). Full details: gateway hook trust model.

​
Supply-chain advisory checking

Mibyan ships with a built-in advisory scanner that flags Python packages in the active venv that match a curated catalog of known-compromised versions (supply-chain worms like the May 2026 mistralai 2.4.6 poisoning). Implementation lives in mibyan_cli/security_advisories.py. How it runs:
  • CLI startup banner. A one-line warning is printed if any advisory matches, with a pointer to mibyan doctor for the full remediation.
  • mibyan doctor. Surfaces every active advisory with version specifics and 2-4 step remediation instructions.
  • Gateway startup. Logged to gateway.log; the first interactive message gets a short operator banner.
Each advisory carries a stable id. Once you have read and acted on it you can dismiss it for good:
The ack is persisted to config.security.acked_advisories and survives restart. Old advisories are intentionally not removed from the catalog — leaving them in place keeps fresh installs warned about historically poisoned versions that might still be cached in a private mirror. The check itself is stdlib-only and runs from one importlib.metadata.version() lookup per advisory, so it’s safe to run on every startup.

​
Lazy install of optional dependencies

PM manages optional Python features as extras from pyproject.toml. Source installers select the all extra. Native bundles include all extras supported by their target. These are different feature sets. When a backend requests an unavailable extra, pm.ensure_import("extra-name") uses the same dependency transaction as plugin admission:
  1. PM checks platform support and security.allow_lazy_installs.
  2. PM prepares a complete environment with the existing extras and enabled plugin requirements.
  3. Without plugin members, it uses the committed lock unchanged. With members, it resolves from the previous selection before a frozen workspace sync.
  4. It validates the candidate before publishing its selection. A failed candidate leaves the previous environment selected.
  5. If the current process uses the previous environment, PM reports that Mibyan must restart. It does not replace imported libraries in place.
Shipped source, locks, and signed payloads remain unchanged. Additional tools and Python environments use writable storage outside the base artifact. Plugin dependencies share the complete environment; they are not isolated Python sandboxes. Compatible transitive dependencies can change, but declared constraints and exact pins remain binding. To disable on-demand installations, run:
Already installed dependencies remain usable. Explicit PM install commands are separate from on-demand installation. A bundle’s frozen feature list, when present with lazy installs disabled, restricts requested Python extra names. This setting is not a blanket ban on explicit plugin admission or manual package-manager commands. The official Docker image also disables on-demand installs through its internal environment policy. For missing dependencies, use mibyan tools and mibyan doctor to identify the requirement. Do not run pip against a signed payload or the system Python. See Package management for installation ownership, diagnostics, and command boundaries.