Cicada.OS

Arch Linux laptop OS with hardened defaults, a launcher monopoly, per-app scopes, and Helium as the browser. Hyprland desk (dock + clickable Wi‑Fi).

Not GrapheneOS. Not AOSP. Graphene’s permission model inspired the scopes sheet; the code is Arch packages + Cicada wrappers (cicada-run, bwrap), which is weaker than Android UIDs. How much boot trust you get is set by hardware — see Hardware tiers.

Pre-alpha: Download ISO · Site · Install

Site: site/ — download / install / features (GitHub Pages). Method: Arch is the package engine. Cicada owns defaults, channel, shell, and profiles.

---

Design rule

Everything here follows one rule, and most of the bugs found so far were violations of it:

Fail closed on anything that would leak a secret. Degrade visibly on anything that only reduces protection.

Knowing the time leaks nothing, so a network that blocks authenticated NTP gets you unauthenticated time and a yellow line in cicada-status — not a laptop that cannot validate a certificate. A VPN kill switch, by contrast, tears the tunnel down rather than let one packet out.

A protection users switch off protects nobody. Defaults are chosen to survive contact with campus Wi-Fi.

---

Hardware tiers

The software is identical across tiers. What changes is which root of trust exists. cicada-hw-trust reports the tier a machine is actually on.

Tier	Hardware	Unlock secret	Evil maid
0	Apple EFI Macs, pre-2016 x86	80–100 bit passphrase, no rate limiting	undetectable
1	TPM2, vendor Secure Boot	short PIN, throttled in hardware	detectable
2	TPM2 + your own SB keys + signed UKI	short PIN + PCR 0+7+11	prevented (boot chain)
3	+ Heads / PureBoot / coreboot	same, on firmware you can build	vendor trust removed

The inversion is the important part. On Tier 0 the passphrase must be long because nothing can rate-limit an offline attack on a disk image. On Tier 1+ the TPM checks the PIN itself and its dictionary-attack lockout counts failures in silicon, so a short PIN is genuinely safe — the same property that makes a 6-digit PIN safe on a Pixel.

Tier 2 is the strongest commodity-x86 story Cicada aims for (your keys + PCR binding). Still missing versus a Pixel: a discrete secure element, and per-app hardware attestation. Do not call that “GrapheneOS.”

Full detail: docs/THREAT_MODEL.md.

---

What it does

Disk & memory

• LUKS2 + Argon2id, memory cost scaled to the machine (¼ RAM, capped 1 GiB)
• cicada-install generates an 80/100-bit passphrase rather than accepting a short one; self-chosen needs 6+ words or 24+ chars
• TPM2 sealing behind a PIN (cicada-tpm-enroll) — never PIN-less, which would make the disk unlock itself on power-on
• USB unlock token as a second factor (cicada-keyfile-enroll), --and mode requires two tokens so a lost stick cannot destroy the disk
• LUKS header backup/verify/restore (cicada-luks-header) — refuses to write to the disk it protects, and says out loud that the backup also undoes the duress and attempt-cap wipes
• No disk swap ever; zram (RAM-only, no writeback) for memory relief
• Core dumps refused at both systemd and kernel level
• noatime, free-RAM scrub at shutdown

Boot chain

• Unified kernel images (cicada-uki) — kernel, initramfs, cmdline and os-release in one signed PE. Without this, Secure Boot signs vmlinuz and nothing else: an initramfs is a cpio and a loader entry is a text file, so an evil maid edits options … init=/bin/sh on the plaintext ESP and the firmware boots it, having verified only the component nobody needed to touch
• That is also what creates PCR 11, so cicada-tpm-enroll --pcrs 0+7+11 binds the disk key to this exact kernel+initramfs+cmdline
• Type-1 entries are removed only after every UKI is built and verified to be a real PE carrying .linux/.initrd/.cmdline; cicada-uki restore reverses it
• Your own Secure Boot keys via cicada-sbctl-enroll; kernel upgrades rebuild and re-sign the UKI through a pacman hook
• Releases are GPG-signed — a published sha256 authenticates nothing, since whoever can swap the ISO can swap the hash beside it

Anti-forensics (mapped against The Law Enforcement and Forensic Examiner's Introduction to Linux, which is the actual playbook)

• Shell history to /dev/null, journal Storage=volatile
• Thumbnail cache and recently-used.xbel wiped at boot
• Amnesic live mode: overlay in RAM, USB yank triggers panic reboot

AFU / seizure

• Suspend and hibernation disabled — a suspended laptop holds the volume key in DRAM indefinitely
• Lid close locks immediately; BFU reboot after 20 min closed, 30 min open
• Hardware watchdog enforces that window even if userspace is compromised; a systemd timer cannot survive an attacker with code execution
• USB authorized_default=0 while locked — the kernel refusing new devices, not userspace policy. Already-authorized devices keep working, so the built-in keyboard survives
• Duress credential at both the LUKS prompt and the lock screen — coercion usually happens to a machine that is already running

Network

• nftables default-deny inbound; VPN kill switch with no ct established exemption in the output chain
• MAC randomization, and DHCP hostname/client-id/DUID withheld so a stable identifier does not undo it
• Quad9 DoT system-wide + DoH in the browser, with ECH; degrades DoH→DoT→plain
• Tor as a first-class cicada-run scope — a network namespace whose only route is Tor, enforced by the kernel, not by a proxy setting an app can ignore
• chrony with NTS from three jurisdictions

Sandboxing

• cicada-run default-denies network, files, camera, mic, USB, sensors
• bubblewrap + xdg-dbus-proxy, per-app scopes
• A seccomp filter, not just namespaces. bubblewrap installs no syscall filter unless handed one, and a namespace hides resources without shrinking the kernel API behind them. cicada-seccomp-gen.sh builds the program at boot from the running kernel's own syscall table — keyring, userfaultfd, perf_event_open, process_vm_readv, open_by_handle_at, pidfd_getfd and the rest denied with EPERM, never a kill
• --new-session (no TIOCSTI back into the launching terminal) and --unshare-ipc (no SysV shared memory between scopes)
• hardened_malloc preloaded, linux-hardened available as a boot entry
• Browser policy is managed (non-overridable): JIT off, WebGPU off, site isolation, post-quantum key agreement

Tamper evidence

• Prevention needs a root of trust; detection only needs somewhere the attacker does not control. cicada-beacon signs a hash over every file on the EFI system partition that decides what the machine boots and pushes it to one device paired by hand (cicada-link) — over LoRa, a USB stick, or a QR code on the screen. On Tier 0, where /boot is plaintext and unmeasured, this is the only measurement of the boot chain that exists
• The witness compares against the last statement it saw: a boot hash that moved on a boot with no kernel upgrade, a seal log whose sequence went backwards, a hardware tier that dropped. It carries no messages, has no account or server, and cannot receive — a witness that could talk back would be a remote-administration channel on a machine designed to have none
• No transport delivered it → exit 2, never a fake success. Someone under coercion acts on the belief that the signal went out

Messaging

• Cicada does not ship a messenger, and will not: the hard parts of Signal are the network and the metadata, not the ratchet, and Matrix's exposure is in the federation design rather than in any client
• What it does instead is host one honestly. cicada-comms bind moves a message store into an encrypted profile, so cicada-lock freezing that profile puts the history at rest — the thing Electron's safeStorage cannot do for itself on a session with no keyring, where it falls back to a constant key compiled into the binary
• cicada-comms doctor reports which of those is true on your machine. shred erases LUKS keyslots when it can and says plainly that it was only an unlink when it cannot

Attack surface

• archiso's installer-ISO services carved out: hypervisor guest agents (a host→guest control channel by design), cloud-init, ModemManager, mirror fetch, pcscd
• 8 setuid binaries stripped, re-applied by a pacman hook after every upgrade
• vivid (CVE-2019-18683) and unused filesystem parsers blacklisted
• Cicada's own root daemons are confined: every unit runs with NoNewPrivileges, a named capability set (the watchdog holds CAP_SYS_BOOT and nothing else; the filter generator holds none), and a syscall filter where one cannot silently break an emergency path. The exceptions are written down in the units themselves with the reason
• AppArmor is put in the kernel's LSM stack (lsm= on the command line). Arch's stock kernel compiles it in but leaves it out of the default list, so enabling the service alone yields an LSM that enforces nothing

---

Tools

Command	Does
cicada-status	The posture in force — time, DNS, MAC, VPN, Tor, swap, core dumps
cicada-verify	Full boot checklist, ordered so failures explain the ones below
cicada-hw-trust	Which tier this machine is on, and the next command to climb one
cicada-tor	status / check / bridges / limits
cicada-portal	Time-boxed plaintext DNS for a captive portal; self-reverting
cicada-logs	Seal log + journal; --verify checks the chain
cicada-run	Launch an app in its scope
cicada-profile	Work / Personal / Burner compartments; end puts one back at rest
cicada-firstrun	One-time wizard: TPM PIN on Tier 1+, USB token on Tier 0
cicada-wifi-diag	Why there is no Wi-Fi
cicada-link	Pair the one device that witnesses this laptop
cicada-beacon	Signed boot-chain statement out of band; verify on the witness
cicada-comms	Where a messenger's history lives, and what that is worth

Escape hatches, typed at the boot menu (e on the entry): cicada.nomalloc disables hardened_malloc; cicada.nowatchdog disables the watchdog. Every control that can render the machine unusable has one.

---

Build

On Apple Silicon this cross-builds an Intel/x86_64 image:

./scripts/build-iso-docker.sh     # → out/cicada-*.iso

Then flash, boot the Mac with Option, and install to an external SSD first. Details in docs/BUILD.md and docs/INSTALL.md.

The image is 1.90 GiB — one file, under GitHub's 2 GiB per-asset cap, so it downloads as a single ISO you can hand straight to Etcher. It was 2.95 GiB and shipped as .part-00/.part-01, which put a cat command between the user and a bootable stick as step one of the install. Three measured cuts, not guesses:

Cut	Saved	Why it was safe
linux-hardened off the live ISO	645 MiB	Neither live loader entry can boot it — both name vmlinuz-linux. It was a kernel, a 233 MiB initramfs, and a second copy of both inside the embedded efiboot.img. Installed systems still get it.
linux-firmware-nvidia	320 MiB	110 MiB of it sat inside every initramfs. Discrete NVIDIA on Wayland is not the daily driver; the laptop still runs on its iGPU.
linux-firmware-marvell	78 MiB	Enterprise NICs and embedded parts.

Both are one command away (pacman -S linux-firmware-nvidia). What stayed is every firmware a laptop actually boots on: intel, amdgpu, atheros, realtek, mediatek, broadcom, radeon, cirrus.

Every ISO carries /usr/share/cicada/BUILD-ID with the commit and whether the tree was dirty. Check it against git rev-parse --short HEAD before concluding anything from a test — a stale ISO is how a fixed bug gets re-reported.

Installing

cicada-install (or the Etch Cicada icon) writes LUKS2 + btrfs + systemd-boot. Three things it does that it did not before, because "hardened" and "installable by a person who is not us" are not in tension:

• No network required. It copies the system you are already running rather than pacstrap-ing 1.2 GiB from the internet. That is faster, it gives you exactly the image that was tested rather than a fresh resolve of it, and it means a laptop whose Wi-Fi is the thing you need Cicada to fix can still be installed. --source network restores the old behaviour.
• Keyboard, timezone and language are asked for — and the keymap is applied before the passphrase prompt. These were hardcoded to us/UTC/en_US, which quietly meant every non-US owner composed their LUKS passphrase on a layout that did not match their keycaps. The keymap is a security setting on this OS: it decides which bytes the physical keys produce, and a mismatch between install time and boot time makes the disk unopenable by anybody.
• It can keep the OS already on the disk. --partition /dev/sdXN --esp /dev/sdXM installs into one partition, leaves the partition table alone and reuses the existing ESP without formatting it. It deliberately does not resize NTFS or APFS — the vendor tool knows things about a live filesystem that an outside resizer does not, and that is where dual-boot installers destroy data. tests/linux/install-guards.sh asserts each refusal fires for its stated reason.

Tests

tests/preflight.sh   # syntax, seal chain, assemble
tests/here.sh        # ~90 checks: defaults, fail-closed CLIs, boot entries
tests/seam.sh        # live vs installed parity — where most bugs have lived
tests/seal.sh        # hash chain + tamper detection
tests/seccomp.sh     # decodes and simulates the sandbox syscall filter
tests/beacon.sh      # signs a statement, edits a fake ESP, requires the alarm
tests/comms.sh       # the at-rest verdict, and everything cicada-comms refuses
tests/afu.sh         # USB gate, gate restore, watchdog arming, escape hatches
tests/donate.sh      # re-derives the donation address checksums; catches drift
tests/iso-size.sh    # does the built ISO still fit one GitHub asset (2 GiB)
tests/linux.sh       # needs Docker: real kernel — nftables, NTS, Tor, duress,
                     # the unprivileged lock-screen duress path, installer guards
tests/boot-verify.sh # run ON the machine after booting (ships as cicada-verify)

These are largely structural — they prove a config says a thing. They do not prove the kernel accepts it. A clean run is necessary, not sufficient; the last build shipped a broken pacman hook that all four suites passed. seccomp.sh is the exception worth copying: it decodes the emitted BPF program and runs the same instruction machine the kernel does, so it fails on a wrong verdict rather than a missing string. boot-verify.sh then loads the program for real.

---

Status

Pre-alpha. Boots to a Hyprland desktop on an Intel MacBook Air. Much of the hardening above is verified structurally and by loading rulesets on real kernels, but has not been exercised on hardware end to end.

Known-unverified, and why. Two lists, because "unverified" was covering two different things and that hid real bugs.

Still needs the Air, cannot be reproduced anywhere else: hardened_malloc running under Helium, whether a chipset watchdog resets this board, whether the kernel refuses a USB device at authorized_default=0, and whether a LoRa radio carries a beacon.

Verified on a real Linux kernel (tests/linux.sh, a privileged linux/amd64 container): the baseline firewall and the VPN kill switch load and the kernel reports the policies claimed here; an established TCP flow stops the moment the switch arms, which is the ct established claim above tested with packets instead of grep; chrony's shipped config completes NTS-KE with all three operators; Tor reaches Bootstrapped 100%; the onion namespace has one route and no physical interface.

Verified on the hardware itself: Doom — Chocolate Doom 3.1.1 on the Freedoom IWAD — has been played on a real machine. That matters far more than a game running: cicada-doom execs cicada-run org.cicada.doom, so it is the only end-to-end evidence that a demanding GPU, audio and input application works inside a bubblewrap scope with the seccomp filter loaded. Nothing in tests/ can produce that result; tests/tirimid.sh only checks that Doom is wired up, never that it runs.

Verified in simulation (tests/afu.sh, tests/beacon.sh): the USB gate, the gate's restore path, watchdog arming, the escape hatches, and the beacon's signing, wire format and alarm.

Six more defects, found the same way. The pattern held: every one was a control that reported success while doing nothing, and in four cases a test was certifying the broken behaviour.

• Session duress was inert at the lock screen — the room it exists for. The PAM hook ran with seteuid, which gives the caller's effective uid: root under sudo, but the desktop user under hyprlock, which is not setuid on Arch. It could not read the 0400 verifier, let alone erase a keyslot, and it exited 0 — the same exit a wrong guess produces. tests/here.sh asserted the string seteuid was present and concluded the wipe worked. There is now a root handler behind cicada-duress.socket, and tests/linux/duress.sh drives it as an unprivileged uid on a real kernel.
• Granting an app the microphone also gave it the session bus. cicada-run bound the whole of XDG_RUNTIME_DIR for MIC=allow, which includes bus (making the xdg-dbus-proxy filter decorative) and systemd/private — i.e. start any user unit, outside the sandbox. Shipped scopes were unaffected, but the Signal scope invites the user to flip exactly that toggle.
• The D-Bus filter failed open. A fixed sleep 0.15 decided the posture by timing; if the proxy was slow the app silently got the real bus.
• cicada-auth was bypassable. The confirmation lived in cicada-profile while the privilege lived in cicada-profile-helper, which is directly reachable via NOPASSWD sudo: cicada-profile-helper dispose work destroyed a profile with no prompt. Same shape for cicada-usb-gate 1, which could reopen USB enumeration while the screen was still locked.
• The published ISO was unsigned (fixed starting v2026.08.19-beta; this cut ships .sha256.asc + cicada-stable.pub).
• The update channel 404’d because channel-latest stayed a draft under GitHub’s 1000-asset cap (fixed by splitting across channel-latest + channel-latest-2 and publishing both after channel-verify-release.sh).
• 60 packages would have 404’d mid-upgrade. GitHub rewrites : to . in release asset filenames — silently, still returning success — so every package carrying a pacman epoch (ffmpeg-2:9.0.1, flatpak-1:1.18.1, fontconfig-2:2.18.3, 57 more) was served under a name the signed database did not name. The release page showed all 1822 assets present either way. channel-build-repo.sh now drops the colon before repo-add, so the database records the name the mirror actually serves; the epoch still reaches pacman through %VERSION%, which is read from .PKGINFO, not from the filename.

The last three are not code, and no test suite was ever going to catch them.

Splitting that list is not bookkeeping — it found five defects, none of which needed hardware to expose. The session duress credential had never worked at all: pam_exec writes the password without a trailing newline, so read returned non-zero, and the hook exited before it ever computed a hash. Also: the USB gate reported success when it had written nothing; cicada-lock stranded the gate closed if it was killed rather than exiting normally; cicada.nomalloc could not rescue a machine that was already broken; and the beacon's Meshtastic call could hang forever on the duress path.

A clean run is still necessary, not sufficient.

Do not rely on this for anything that matters yet. See docs/ROADMAP.md and docs/test-results.md.

What Cicada does not claim

• It is not GrapheneOS, and on Tier 0 hardware it cannot approach it. No secure element, no verified boot, no key-guessing rate limit.
• Evil maid is unprevented on Tier 0. /boot is plaintext, unsigned and unmeasured; modifying the initramfs to capture the passphrase needs a screwdriver, not an exploit. cicada-beacon makes it detectable by a device the attacker does not hold, which is a different and weaker claim: it tells you afterwards, and only if you paired a witness and read what it says.
• Intel ME / AMD PSP are present on every tier and are not neutralized.
• Tor does not hide that you use Tor without a pluggable transport, and obfs4/snowflake are AUR-only — on a censored network use Tor Browser, which bundles its own.
• Messengers are confined by Flatpak, not by Cicada. Signal, SimpleX and Element run under Flatpak's sandbox with floors written below Flathub's defaults — not under a cicada-run scope, and not under the syscall filter above. That boundary is deliberate and permanent, not a gap (docs/COMMS.md). What Cicada contributes is where the message store lives and what it is worth at rest.
• A seized-unlocked or compromised machine makes everything above moot.

Threat model: docs/THREAT_MODEL.md · docs/ATTACKS.md · docs/GRAPHENE_PARITY.md · docs/SANDBOX.md · docs/DESIGN.md · docs/BEACON.md · docs/COMMS.md