Test evidence and reproduction artifacts for a Fastly Next-Gen WAF agent fail-open condition during Kubernetes pod startup, where traffic passes uninspected until rules load.
Evidence for the writeup Fastly Next-Gen WAF Agent: Fail-Open During Kubernetes Pod Startup.
In the Kubernetes sidecar deployment Fastly publishes for its Next-Gen WAF (formerly Signal Sciences), the NGINX module and sigsci-agent pass traffic uninspected whenever the agent has not loaded a ruleset:
errno=2) and fails open.RPC.PreRequest with WAFResponse 200 and X-SigSci-Agent: disabled, and keeps doing so until a download succeeds.The pod reports Ready throughout. With default settings, the agent never logs that it is disabled. Fastly's opt-in startup probe holds the pod out of service until rules load; Fastly's published Helm values do not include it.
Fastly's position is that this is intended fail-open behavior and not a vulnerability.
| Date | 2026-11-12 (America/Chicago) |
| Agent | signalsciences/sigsci-agent:4.83.2, index digest sha256:4aa2f5296b5f2eeb1a71e79643a93afb19ce85a635a21895ff35b2d0701578d2 |
| Ingress controller | signalsciences/sigsci-nginx-ingress-controller:1.15.1-fastly1.4.0, index digest sha256:a4c66bfe5c0b06c6507ce89881ba891711ab3009b2d5d934db192a91e0f452ba |
| Module | sigsci-module-nginx-native 1.4.0, nginx 1.27.1 |
| Layout | Fastly's sigsci-values.yaml (commit 9794dfb) on ingress-nginx chart 4.15.1 |
| Cluster | kind v1.31.0 on podman, macOS arm64 |
disabled state: 20,304 decoded agent responses, one distinct value.initialDelaySeconds of 10, 30 and 60 s: traffic, including attack payloads, reached the upstream after each delay expired.Full results, with each claim labeled OBSERVED, INFERRED or UNTESTED and tied to an evidence file, are in REPORT.md.
No Fastly credentials were used. The agent could never load a ruleset, so a healthy start was not observed: whether a valid agent ever opens its socket before its first ruleset loads, the startup probe's 503 to 200 transition, and how these requests appear in Fastly's console are all untested. The 60-second window in the egress-blocked runs comes from how egress was blocked, not a typical duration. See section 6 and 9.9 of the report.
Testing was black-box: logs, documented flags, the agent's documented RPC debug dump, and HTTP headers. No binaries were decompiled or disassembled.
| Path | |
|---|---|
REPORT.md | Full report, method, results, limitations, reproduction steps |
manifests/ | Exact Helm values and Kubernetes manifests for each variant |
raw/<run>/ | Per-run evidence: agent, NGINX, echo-server and generator logs, decoded RPC, statsd, pod and endpoint watches |
src/ | Orchestration, analysis and RPC decoding scripts |
SHA256SUMS | Checksums of every evidence file as collected |
The access keys in the manifests are format-valid dummy values (00000000-0000-0000-0000-000000000000 and 43 A characters), not credentials. Upstream sigsci-values.yaml in raw/ is Fastly's file, copied unmodified for reference.
sha256sum -c SHA256SUMS
SHA256SUMS covers the evidence as collected on 2026-11-12 and 2026-11-13. This README was added afterward and is not included in it. REPORT.md was edited for publication on 2026-11-17 (environment notes trimmed), so the check reports that one file as changed; the original is in commit 32930cd. Every file under raw/, manifests/ and src/ is unchanged.
See section 7 of REPORT.md. Images should be pulled by the digests above.