Skip to main content

Polaris 1.0 Reference demo

One always-on deployment, polaris-1-0-reference, that shows prevention, detection and recovery of a kernel attack. Plain Polaris, no Invary.

Unlike the three older demos there is no broker. Everyone who runs this is a member of the demo's GCP project (echo-487317), so the CLI uses their own gcloud login instead. Setup per machine is one command:

gcloud auth login # a member of echo-487317

The CLI lives in the polaris-deploy checkout at scripts/polaris-1-0.sh. It reads the endpoint and tokens from Secret Manager with your own login, so you do not need to have deployed anything.

Pinned version​

Run the demo from the tag polaris-1-0-demo-v2, not from main, so work on main cannot change the demo under you. Keep a checkout of its own at the tag and install the CLI from it:

git -C polaris-deploy fetch --tags
git -C polaris-deploy worktree add ../polaris-deploy-demo polaris-1-0-demo-v2
polaris-deploy-demo/scripts/polaris-1-0.sh install

Polaris Desktop carries the same tag; the demo's TestFlight build is cut from it. To move the demo forward, check that scripts/tests/polaris-1-0-demo-holds.sh passes on main (CI runs it), then tag a new version and point the checkout at it with git -C polaris-deploy-demo checkout <new tag>.

Commands​

CommandWhat it does
./scripts/polaris-1-0.sh watchStreams each attestation as it arrives, appended below the last, so you can scroll back to any earlier one. Ctrl-C stops it. Needs only the read-only token, which the CLI fetches for you.
./scripts/polaris-1-0.sh statusThe board once: per-layer verdicts and learning.
./scripts/polaris-1-0.sh desktopPrint the endpoint and token to paste into Polaris Desktop.
./scripts/polaris-1-0.sh preventAct 1: load a module signed by an untrusted key and watch the kernel refuse it. The verdict stays pass.
./scripts/polaris-1-0.sh attackAct 2: load the trusted-signed rogue module and leave it loaded. The attacker then opens an outbound connection to exfiltrate. L3.1 fails on the module and L3 alerts on the connection in the same cycle, the key is revoked, the workload is cut off.
./scripts/polaris-1-0.sh restoreAct 3: unload the module and watch the verdict return to pass. It also clears the ransomware act.
./scripts/polaris-1-0.sh ransomAct 4: an attacker who logged in with a stolen SSH key drops ransomware. It locks the files on a decoy share and calls home. IMA runtime finds a program it never measured, and the eBPF profile finds a binary the host does not run and a connection to an unrecognized address. It clears itself after 5 minutes, or run restore.

Running the demo​

Two windows, or one: watch appends each attestation and the acts print in the same form, so it works in its own window or in the main one between acts (Ctrl-C to stop it). In the first window, the live stream, with Polaris Desktop open beside it:

./scripts/polaris-1-0.sh watch

For Desktop, run ./scripts/polaris-1-0.sh desktop and paste the endpoint and token into the Security screen (Add Endpoint).

In the second window, the three acts, in order:

./scripts/polaris-1-0.sh prevent
./scripts/polaris-1-0.sh attack
./scripts/polaris-1-0.sh restore
./scripts/polaris-1-0.sh ransom
./scripts/polaris-1-0.sh restore

Act 1 shows that a normal attacker cannot even load the module: the kernel runs with Secure Boot and lockdown=integrity, so it refuses a module no trusted key signed, root or not. Act 2 stands in for an insider with a stolen signing key: the module loads, and Polaris catches the load itself. L3.1 fails with unauthorized kernel module loaded: roguebpf. With the eBPF monitor blinded, the attacker opens an outbound connection to exfiltrate, and L3 network alerts on it in the same cycle (connected to unrecognized destination 192.0.2.1:443, a reserved test address so nothing leaves the sandbox). The workload key is revoked, the proxy closes its gate, the secret lease is dropped and the SPIFFE socket stops issuing. Act 3 unloads the module, stops the connection and everything returns. Pass --no-exfil to run the module-only version.

Act 4 is the attack most people have met: ransomware. Someone logs in with a stolen SSH key and starts it. It encrypts the files on a made-up finance share, leaves a ransom note, and tries to reach 198.51.100.66:443, another reserved test address that the deployment's egress firewall blocks anyway. Polaris does not stop the encryption; it catches the program within one attestation, the binary the host profile never runs and its connection attempt, both naming /var/tmp/.sysupdate/encryptor, and revokes the key so the workload's secrets are cut off. restore stops it and removes the files.

Notes​

  • The attack acts need IAP SSH and sudo on the CVM, which a project owner has. Watching needs only the token.
  • The deployment runs with L3.1 and l3_binaries enforcing, so Act 2 and the ransomware act revoke the key rather than only alerting. A redeploy resets the dials; set them again with scripts/deploy.sh promote --project echo-487317 --name polaris-1-0-reference --layers l31,l3_binaries.
  • The rogue module and both signed copies are baked into the CVM image and only ever run on this throwaway sandbox.