evil-echo

paddock-evil-echo — hostile test harness

⚠️ Do NOT use as a real harness. This image deliberately attempts security-relevant operations to validate Paddock's defences. It exists only for adversarial E2E coverage.

Purpose

paddock-evil-echo is the test-side counterpart to paddock-echo. It runs as a HarnessRun's main container (or as a Workspace seed image) and attempts a sequence of hostile actions specified by CLI flags. Each action emits one line of structured JSON describing the attempt's outcome. E2E tests parse the JSON and assert that Paddock's defences denied the action.

Flags

See docs/internal/security-audits/2026-04-25-v0.4-test-gaps.md §4 for the full catalogue. As of v0.5.0, 15 flags are supported:

• --bypass-proxy-env — unset HTTPS_PROXY before subsequent flags.
• --connect-raw-tcp <host:port> — raw TCP dial.
• --connect-ip-literal <ip:port> — same, but flag name signals IP-literal test.
• --connect-loopback <port> — TCP dial to 127.0.0.1.
• --read-secret-files <glob> — list/stat files matching glob.
• --read-pvc-git-config — read /workspace/.git/config and /workspace/.paddock/repos.json.
• --probe-broker <url> — POST a synthetic bearer to the broker.
• --probe-imds — connect to 169.254.169.254:80 (cloud metadata service).
• --probe-env-override — report the values of HTTPS_PROXY etc.
• --probe-provider-substitution-host <url> — request to non-allowlisted host with a synthetic broker bearer.
• --exfil-via-dns <host> — DNS lookup to host.
• --read-other-tenant-pvc <namespace> — attempt cross-namespace PVC read.
• --forge-ca-cert <fqdn> — read per-run CA private key.
• --flood-connect-raw-tcp <host:port> — 50 sequential TCP dials (DoS probe).
• --smuggle-headers <name=value> — request with extra header through proxy.

Build

make image-evil-echo

Image is tagged paddock-evil-echo:dev. Not pushed to public registries by default.

Output format

One JSON object per flag, one line per object:

{"flag": "--connect-raw-tcp", "target": "10.0.0.1:443", "result": "denied", "error": "i/o timeout"}

result is one of success, denied, error, skipped. success means the attack succeeded (Paddock's defence failed); E2E tests typically assert denied or error.

Exit code

Always 0 when at least one flag ran (regardless of attack success). Success/failure is in the JSON, not the exit code, so a hostile run that "successfully" fails is still detectable by the test harness.