Before this fix, RunScan declared `scanErrors []string` but never
appended to it. As a result:
- the summary Info log ("network target scan completed") always
reported `"errors": 0`, regardless of how many endpoints failed
- the DiscoveryReport's `Errors` field — stored on the scan record
and surfaced in the GUI scan history — was always nil
Operators who needed to understand scan failures had to enable Debug
logging and grep through the noise of expected sweep-scan connection
refusals. The per-endpoint log level (Debug) is deliberate and correct
— scanning a /24 typically produces 200+ connection-refused results,
and logging each at Warn would create massive log spam at default
verbosity. The bug was the silent loss of the aggregate count.
This commit:
- extracts the partitioning logic into `collectScanResults`, a pure
method that splits per-endpoint results into discovered certificate
entries and a list of endpoint error strings
- populates the errors list with "<address>: <error>" so the scan
record correlates failures back to specific endpoints
- preserves the existing Debug-level per-endpoint log (sweep noise
discipline) — no change to default-verbosity log output
The summary Info log's "errors" field and the DiscoveryReport's Errors
field now reflect the true failure count. Debug detail remains
available for operators diagnosing specific endpoints.
Audit scope note: the M-9 finding narrative implied broad Debug-level
hiding of real errors across AWS SM, Azure KV, GCP SM, and network
scan sentinel agents. On investigation, the three cloud-discovery
connectors (awssm, azurekv, gcpsm) already use appropriate Warn/Error
discipline for per-item and root-level failures. Only the network
scanner had a silent observability gap, and it was a missed append
rather than a misapplied log level. See audit resolution log for
full details.
CWE: CWE-778 (Insufficient Logging) — aggregate failure count lost.
Tests: 4 new unit tests on collectScanResults covering the
aggregation path (success + failure mix), all-success, all-failed,
and empty-input degenerate cases. All tests pass with -race.
Verification:
- go build ./cmd/server/... ./cmd/agent/... ./cmd/mcp-server/... ./cmd/cli/... exit 0
- go vet ./... exit 0
- go test -race -count=1 -timeout 300s [full CI race path] exit 0
- golangci-lint run ./... --timeout 5m (v2.11.4) 0 issues
- govulncheck ./... (@latest) 0 in-code vulnerabilities
- go test -count=1 -cover ./internal/service/... 68.0% (> 55% threshold)
Invariants preserved:
- collectScanResults signature: method on *NetworkScanService,
input []domain.NetworkScanResult, return ([]DiscoveredCertEntry, []string)
- Debug log key names unchanged ("address", "error")
- DiscoveryReport schema unchanged (Errors field already existed)
- Sentinel agent ID "server-scanner" unchanged
- No migration, no API, no wire-format change
Refs: M-9 Medium finding; audit resolution log appended in follow-up
commit on workspace-level audit report.
The webhook notifier would previously accept any operator-configured URL
and hand it to http.Client without validation. That exposed two
SSRF classes (CWE-918):
* Reserved-address reachability — a misconfigured or adversarial
webhook URL pointing at 127.0.0.1, ::1, 169.254.169.254 (cloud
metadata), or 0.0.0.0 would succeed, exfiltrating request bodies
to local services or leaking short-lived cloud credentials.
* DNS rebinding — a hostname resolving to a public IP at validation
time and to a reserved IP at dial time would bypass any
URL-string-only check.
Fix installs two independent layers:
* validation.ValidateSafeURL runs at config-ingest time and before
every outbound POST. It rejects non-HTTP(S) schemes, empty hosts,
and literal reserved-IP hosts with a clear operator-facing error.
This is a fast early diagnostic.
* validation.SafeHTTPDialContext is installed on the webhook
http.Transport. It re-resolves the host at dial time, rejects any
resolved address whose address lies in a reserved range (loopback,
link-local, multicast, broadcast, unspecified, IPv6
link-local/multicast), and pins the resolved IP into the final
dial address so the TLS handshake targets the exact IP the guard
approved. This is the authoritative, TOCTOU-safe defence against
DNS rebinding.
The two layers are complementary — validateURL fails fast on obvious
misconfiguration; SafeHTTPDialContext fails closed when DNS changes
between validation and dial.
The existing unexported isReservedIP helper in
internal/service/network_scan.go is extracted into
internal/validation.IsReservedIP with byte-identical behaviour so the
webhook notifier and the network scanner share a single authoritative
reserved-address list. RFC 1918 ranges remain intentionally allowed
(certctl's self-hosted design). Broader unspecified / IPv6 link-local
coverage lives only in the stricter dial-time policy, where it belongs
for outbound HTTP egress.
Test seam: Connector gains an unexported validateURL func field and a
same-package newForTest constructor that installs a permissive
validator and the stdlib default transport. Production callers cannot
reach this constructor because it is unexported; only same-package
tests (package webhook) can use it. Same-package happy-path tests call
newForTest so they can point at httptest loopback servers without
being blocked by the production guard. The four SSRF-rejection tests
that verify the guard itself still call New so they exercise the real,
strict validator. This keeps the production SSRF defence
unconditionally on in real code while preserving legitimate unit-test
coverage.
Tests
-----
* internal/validation/ssrf_test.go (new) — 16-subtest pin on
IsReservedIP that is byte-identical with the original network-
scanner behaviour; ValidateSafeURL accept/reject matrix covering
HTTPS/HTTP, reserved-literal IPv4/IPv6, dangerous schemes
(file/gopher/ftp/javascript/data/ldap/dict/jar), missing hosts,
and malformed inputs; SafeHTTPDialContext rejects literal reserved
addresses and hosts resolving to reserved addresses (DNS-rebinding
coverage via localhost).
* internal/connector/notifier/webhook/webhook_test.go — happy-path
tests switched to newForTest; production-guard SSRF-rejection
tests (TestValidateConfig_RejectsReservedURLs,
TestValidateConfig_RejectsDangerousScheme,
TestPostWebhook_RejectsReservedURL,
TestPostWebhook_RejectsDangerousScheme) continue to call New so
they exercise the unconditionally-installed production validator.
Wire-format invariants preserved
--------------------------------
* Outbound HTTP request shape (method, headers, body, HMAC
signature) unchanged.
* network_scan.go behaviour unchanged — validation.IsReservedIP is
byte-identical with the deleted helper.
* RFC 1918 (10/8, 172.16/12, 192.168/16) remain allowed for both
outbound webhook and CIDR expansion, matching the self-hosted
design.
Verification
------------
* go test -race ./internal/validation/... ./internal/connector/
notifier/webhook/... ./internal/service/... — green.
* Full-suite go test -race ./... — green (GOTMPDIR=/dev/shm to
sidestep full /tmp on the sandbox host).
* Coverage gates pass: service 68.8% >= 55%, handler 83.6% >= 60%,
domain 82.0% >= 40%, middleware 63.8% >= 30%. Overall 67.8%.
Webhook package 91.5% line coverage; validation package
ValidateSafeURL/SafeHTTPDialContext 78-100% per function.
* govulncheck ./... — no vulnerabilities found.
* golangci-lint run on touched H-4 production code — clean. Pre-
existing errcheck/gosimple warnings in scope-adjacent files
(webhook_test.go:270 w.Write, network_scan.go:120/173/265/305)
verified against 3853b74 to predate this commit; left alone per
scope guard.
Operational notes
-----------------
* No migration needed. The guard is pure Go code; existing webhook
configs continue to work unless they point at reserved addresses,
in which case they now fail closed with a clear error.
* Existing operators who rely on webhook POST to 127.0.0.1 or
::1 (e.g., local receivers on the same host as certctl-server)
must expose their receiver on an RFC 1918 address or public IP.
This is deliberate — the threat model for webhook notifiers
includes untrusted operator-supplied URLs.
Scope guard: H-4 only. H-5, H-6, M-*, L-*, and I-* findings remain
open and are tracked separately. No drive-by refactors.
Three bugs fixed:
- Docker Compose only mounted migration 000001; migrations 000002-000007
(profiles, agent groups, revocation, discovery, network scans) never ran,
breaking half the demo features. Now mounts all 7 migrations in order.
- Network Scans page crashed with pq.Array scan error because lib/pq
doesn't support []int, only []int64. Changed Ports field accordingly.
- Dashboard pie chart displayed "RenewalInProgress" without spaces.
Added formatStatus() helper for PascalCase → spaced display.
Also adds first-run demo experience improvements:
- 9 discovered certificates (filesystem + network scan mix)
- 3 discovery scans with recent timestamps
- 2 AwaitingApproval renewal jobs for approval workflow demo
- CERTCTL_NETWORK_SCAN_ENABLED=true in Docker Compose
- Network scan targets seeded with last_scan results
- Version badge updated to v2.0.5
- Docs updated (quickstart, advanced demo) to reference seeded data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
M21 adds server-side active TLS scanning of CIDR ranges with concurrent
probing, sentinel agent pattern for pipeline reuse, and full CRUD API for
scan targets. M22 adds Prometheus exposition format endpoint alongside
existing JSON metrics. Comprehensive documentation audit updates all docs
to reflect 91 endpoints, 19 tables, 6 scheduler loops, and 900+ tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
shankar0123