Enforce Agent Rules
Enforcement is level 2 of SilentShield's AI Agent Control: instead of only recording AI agent visits, your server or edge actively applies your per-key policy — each agent is allowed, throttled, or blocked. It builds directly on observe mode. The enforcement components (Next.js middleware, sidecar binary, Cloudflare Worker) are currently provided on request during early access — contact [email protected].
Observe First, Then Enforce
Our clear recommendation: enable observation first, let it run for a week or two, and review the report before you block anything — the same staged rollout as Wordfence's learning mode or DMARC (none → quarantine → reject). Only the report shows which AI services actually visit your site and what a block would cost you, for example citations in AI search. Enforcing without that picture means deciding blind.
How Enforcement Works
Every enforcer shares the same contract: it fetches a policy bundle for your API key, verifies the bundle's Ed25519 signature against SilentShield's pinned public keys, and then decides locally — allow, deny, or throttle — without any call on the request path.
You fetch the pinned signing keys once from https://api.silentshield.io/.well-known/silentshield-agent-keys. And every enforcer fails open: on any error — network, signature, expired bundle — the request is let through. Enforcement can never take your site down.
Setups
Go (net/http)
A self-contained enforcer — standard library only, no SDK. It fetches and verifies the signed policy in the background (Ed25519, pinned keys) and decides per request; wrap your handler with its middleware. Fail-open: any error or monitor mode lets the request through.
// Self-contained SilentShield agent-policy enforcer — stdlib only, no SDK.
// Drop this in, wrap your handler with New(cfg).Middleware, and blocked bots
// get a 403 (throttled ones a 429). Fail-open by design: any error, an
// unverifiable bundle, or monitor mode lets the request through — enforcement
// can never take your site down.
package ssenforce
import (
"bytes"
"crypto/ed25519"
"encoding/base64"
"encoding/json"
"io"
"net"
"net/http"
"strconv"
"strings"
"sync"
"time"
)
// Config for the enforcer. PolicyURL + APIKey + at least one trusted key are
// required; with any missing the enforcer is a no-op (always allow).
type Config struct {
PolicyURL string // https://api.silentshield.io/api/v1/agent/policy
KeysURL string // https://api.silentshield.io/.well-known/silentshield-agent-keys
APIKey string // your publishable site key (x-api-key)
}
type signedBundle struct {
Payload string `json:"payload"`
Alg string `json:"alg"`
Kid string `json:"kid"`
Sig string `json:"sig"`
}
type bundle struct {
FormatVersion int `json:"format_version"`
Mode string `json:"mode"`
QuotaEnabled bool `json:"quota_enabled"`
Rules []rule `json:"rules"`
Agents []agent `json:"agents"`
}
type agent struct {
Slug string `json:"slug"`
Category string `json:"category"`
UATokens []string `json:"ua_tokens"`
CIDRs []string `json:"cidrs"`
}
type rule struct {
Match match `json:"match"`
PathPattern string `json:"path_pattern"`
Methods []string `json:"methods"`
Action string `json:"action"`
RateLimit *rateLimit `json:"rate_limit"`
}
type match struct {
Type string `json:"type"`
Value string `json:"value"`
}
type rateLimit struct {
Requests int `json:"requests"`
WindowSec int `json:"window_sec"`
}
// Enforcer fetches, verifies and caches the signed policy, refreshing every
// 5 minutes in the background, and decides per request.
type Enforcer struct {
cfg Config
enabled bool
mu sync.RWMutex
b *bundle
cidrs map[string][]*net.IPNet
countersMu sync.Mutex
counters map[string]*window
}
type window struct {
bucket int64
count int
}
// New builds an enforcer and starts background refresh. Call once at startup.
func New(cfg Config) *Enforcer {
e := &Enforcer{cfg: cfg, counters: map[string]*window{}}
e.enabled = cfg.PolicyURL != "" && cfg.KeysURL != "" && cfg.APIKey != ""
if !e.enabled {
return e
}
_ = e.refresh() // best-effort initial load
go func() {
t := time.NewTicker(5 * time.Minute)
defer t.Stop()
for range t.C {
_ = e.refresh()
}
}()
return e
}
// Middleware wraps next, blocking disallowed bots. Monitor mode / fail-open
// never block.
func (e *Enforcer) Middleware(next http.Handler) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
// Optional: exempt your own trusted traffic (e.g. logged-in users) from
// enforcement. Plug in your own check — it runs first, so a matching
// request is never blocked. Keep it cheap (inspect your session cookie /
// JWT), no I/O.
// if isLoggedIn(r) {
// next.ServeHTTP(w, r)
// return
// }
if block, status, retry := e.decide(r); block {
if retry > 0 {
w.Header().Set("Retry-After", strconv.Itoa(retry))
}
w.WriteHeader(status)
reportBlock(e.cfg.APIKey, r, status) // feed the dashboard's blocked-bots report
return
}
next.ServeHTTP(w, r)
})
}
// reportBlock fire-and-forgets a report that this request was blocked, so the
// SilentShield dashboard's "blocked bots" report has data. Best-effort: its own
// goroutine + timeout, all errors ignored — it never delays the response.
func reportBlock(apiKey string, r *http.Request, status int) {
outcome := "deny"
if status == http.StatusTooManyRequests {
outcome = "throttle"
}
s := map[string]any{"ua": r.UserAgent(), "ip": remoteIP(r), "path": r.URL.Path, "method": r.Method, "outcome": outcome}
body, _ := json.Marshal(map[string]any{"sightings": []any{s}})
go func() {
req, err := http.NewRequest("POST", "https://api.silentshield.io/api/v1/agent/telemetry", bytes.NewReader(body))
if err != nil {
return
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("x-api-key", apiKey)
if resp, err := (&http.Client{Timeout: 3 * time.Second}).Do(req); err == nil {
resp.Body.Close()
}
}()
}
func (e *Enforcer) decide(r *http.Request) (block bool, status, retry int) {
if !e.enabled {
return false, 0, 0
}
e.mu.RLock()
b, cidrs := e.b, e.cidrs
e.mu.RUnlock()
if b == nil || b.Mode != "enforce" {
return false, 0, 0 // no bundle or monitor → never block
}
slug, category := "", ""
best := 0
ua := strings.ToLower(r.UserAgent())
for _, a := range b.Agents {
for _, tok := range a.UATokens {
tl := strings.ToLower(tok)
if tl != "" && strings.Contains(ua, tl) && len(tl) > best {
best, slug, category = len(tl), a.Slug, a.Category
}
}
}
verified := false
if slug != "" {
ip := net.ParseIP(remoteIP(r))
for _, n := range cidrs[slug] {
if ip != nil && n.Contains(ip) {
verified = true
break
}
}
}
ru := evaluate(b.Rules, slug, category, verified, r.URL.Path, r.Method)
if ru == nil {
return false, 0, 0
}
switch ru.Action {
case "deny":
return true, http.StatusForbidden, 0
case "throttle":
if b.QuotaEnabled && ru.RateLimit != nil {
key := slug
if key == "" {
key = "cat:" + category
}
if ok, ra := e.allowThrottle(key, ru.RateLimit); !ok {
return true, http.StatusTooManyRequests, ra
}
}
}
return false, 0, 0
}
// evaluate walks the rules in order, first match wins (nil → allow).
func evaluate(rules []rule, slug, category string, verified bool, path, method string) *rule {
for i := range rules {
ru := &rules[i]
if !matchApplies(ru.Match, slug, category, verified) {
continue
}
if !pathMatches(ru.PathPattern, path) {
continue
}
if !methodMatches(ru.Methods, method) {
continue
}
return ru
}
return nil
}
func matchApplies(m match, slug, category string, verified bool) bool {
switch m.Type {
case "any":
return true
case "agent_slug":
return slug != "" && slug == m.Value
case "agent_category":
return category != "" && category == m.Value
case "unsigned":
return !verified
}
return false
}
func pathMatches(pattern, path string) bool {
if pattern == "" || pattern == "*" {
return true
}
if strings.HasSuffix(pattern, "*") {
return strings.HasPrefix(path, strings.TrimSuffix(pattern, "*"))
}
return path == pattern
}
func methodMatches(methods []string, method string) bool {
if len(methods) == 0 {
return true
}
for _, m := range methods {
if strings.EqualFold(strings.TrimSpace(m), method) {
return true
}
}
return false
}
func (e *Enforcer) allowThrottle(key string, rl *rateLimit) (bool, int) {
if rl.Requests <= 0 || rl.WindowSec <= 0 {
return true, 0
}
now := time.Now().Unix()
bucket := now / int64(rl.WindowSec)
e.countersMu.Lock()
defer e.countersMu.Unlock()
c := e.counters[key]
if c == nil || c.bucket != bucket {
c = &window{bucket: bucket}
e.counters[key] = c
}
c.count++
if c.count > rl.Requests {
retry := int((bucket+1)*int64(rl.WindowSec) - now)
if retry < 1 {
retry = 1
}
return false, retry
}
return true, 0
}
// refresh fetches + verifies the bundle and swaps it in. On any error the
// previous bundle is kept (fail-open).
func (e *Enforcer) refresh() error {
keys, err := e.fetchKeys()
if err != nil {
return err
}
req, _ := http.NewRequest(http.MethodGet, e.cfg.PolicyURL, nil)
req.Header.Set("x-api-key", e.cfg.APIKey)
resp, err := http.DefaultClient.Do(req)
if err != nil {
return err
}
defer resp.Body.Close()
if resp.StatusCode != http.StatusOK {
return nil // 304/anything else → keep last good bundle
}
raw, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
if err != nil {
return err
}
b, err := verifyBundle(raw, keys)
if err != nil {
return err
}
cidrs := map[string][]*net.IPNet{}
for _, a := range b.Agents {
for _, c := range a.CIDRs {
if _, n, err := net.ParseCIDR(c); err == nil {
cidrs[a.Slug] = append(cidrs[a.Slug], n)
}
}
}
e.mu.Lock()
e.b, e.cidrs = b, cidrs
e.mu.Unlock()
return nil
}
func (e *Enforcer) fetchKeys() (map[string]ed25519.PublicKey, error) {
resp, err := http.Get(e.cfg.KeysURL)
if err != nil {
return nil, err
}
defer resp.Body.Close()
var payload struct {
Keys []struct {
Kid string `json:"kid"`
Key string `json:"key"`
} `json:"keys"`
}
if err := json.NewDecoder(resp.Body).Decode(&payload); err != nil {
return nil, err
}
out := map[string]ed25519.PublicKey{}
for _, k := range payload.Keys {
if raw, err := base64.StdEncoding.DecodeString(k.Key); err == nil && len(raw) == ed25519.PublicKeySize {
out[k.Kid] = ed25519.PublicKey(raw)
}
}
return out, nil
}
// verifyBundle checks the Ed25519 signature over the raw payload and returns the
// decoded bundle. The signed bytes are the payload JSON exactly (no hashing).
func verifyBundle(body []byte, keys map[string]ed25519.PublicKey) (*bundle, error) {
var sb signedBundle
if err := json.Unmarshal(body, &sb); err != nil {
return nil, err
}
if sb.Alg != "ed25519" {
return nil, errInvalid
}
pub, ok := keys[sb.Kid]
if !ok {
return nil, errInvalid
}
payload, err := base64.StdEncoding.DecodeString(sb.Payload)
if err != nil {
return nil, err
}
sig, err := base64.StdEncoding.DecodeString(sb.Sig)
if err != nil {
return nil, err
}
if !ed25519.Verify(pub, payload, sig) {
return nil, errInvalid
}
var b bundle
if err := json.Unmarshal(payload, &b); err != nil {
return nil, err
}
if b.FormatVersion > 1 {
return nil, errInvalid // a newer wire format we can't be sure we understand
}
return &b, nil
}
// remoteIP is the source IP for CIDR verification. Behind a trusted proxy,
// restore the real client IP here (do NOT trust a spoofable header for the
// verification decision).
func remoteIP(r *http.Request) string {
ip, _, err := net.SplitHostPort(r.RemoteAddr)
if err != nil {
return r.RemoteAddr
}
return ip
}
type sentinel string
func (s sentinel) Error() string { return string(s) }
const errInvalid = sentinel("invalid bundle")
// Wiring:
// enf := ssenforce.New(ssenforce.Config{
// PolicyURL: "https://api.silentshield.io/api/v1/agent/policy",
// KeysURL: "https://api.silentshield.io/.well-known/silentshield-agent-keys",
// APIKey: "YOUR_API_KEY",
// })
// http.ListenAndServe(":8080", enf.Middleware(mux))Node.js / Express
The same enforcer for Node.js, using only the built-in crypto module. Register it as the first middleware so it runs before your routes. Fail-open by design.
// Self-contained SilentShield agent-policy enforcer for Node.js / Express —
// no SDK, only the built-in `crypto`. Wrap your app with enforcer(cfg) and
// blocked bots get a 403 (throttled ones a 429). Fail-open by design: any
// error, an unverifiable bundle, or monitor mode lets the request through.
const crypto = require("crypto");
// Raw 32-byte Ed25519 public key → a verifiable KeyObject (DER SPKI wrap).
const ED25519_SPKI_PREFIX = Buffer.from("302a300506032b6570032100", "hex");
function toPublicKey(rawBase64) {
const raw = Buffer.from(rawBase64, "base64");
if (raw.length !== 32) return null;
return crypto.createPublicKey({
key: Buffer.concat([ED25519_SPKI_PREFIX, raw]),
format: "der",
type: "spki",
});
}
function ipInCidr(ip, cidr) {
// Node's net has no CIDR test; do a byte-mask compare on the parsed buffers.
const [subnet, bitsStr] = cidr.split("/");
const bits = parseInt(bitsStr, 10);
const a = ipToBuf(ip);
const b = ipToBuf(subnet);
if (!a || !b || a.length !== b.length || bits < 0 || bits > a.length * 8) return false;
const full = Math.floor(bits / 8);
for (let i = 0; i < full; i++) if (a[i] !== b[i]) return false;
const rem = bits % 8;
if (rem === 0) return true;
const mask = (0xff << (8 - rem)) & 0xff;
return (a[full] & mask) === (b[full] & mask);
}
function ipToBuf(ip) {
if (net_isIPv4(ip)) return Buffer.from(ip.split(".").map((n) => parseInt(n, 10)));
try {
// Expand IPv6 to 16 bytes via the built-in parser.
const parts = require("net").isIPv6(ip) ? ip : null;
if (!parts) return null;
return ipv6ToBuf(ip);
} catch {
return null;
}
}
function net_isIPv4(ip) {
return require("net").isIPv4(ip);
}
function ipv6ToBuf(ip) {
// Minimal IPv6 → 16-byte buffer (handles "::" compression).
let [head, tail] = ip.split("::");
const h = head ? head.split(":") : [];
const t = tail ? tail.split(":") : [];
const missing = 8 - (h.length + t.length);
if (missing < 0) return null;
const groups = [...h, ...Array(missing).fill("0"), ...t];
const buf = Buffer.alloc(16);
for (let i = 0; i < 8; i++) {
const v = parseInt(groups[i] || "0", 16);
buf.writeUInt16BE(v, i * 2);
}
return buf;
}
function matchApplies(m, slug, category, verified) {
switch (m && m.type) {
case "any": return true;
case "agent_slug": return slug !== "" && slug === m.value;
case "agent_category": return category !== "" && category === m.value;
case "unsigned": return !verified;
default: return false;
}
}
function pathMatches(pattern, path) {
if (!pattern || pattern === "*") return true;
if (pattern.endsWith("*")) return path.startsWith(pattern.slice(0, -1));
return path === pattern;
}
function methodMatches(methods, method) {
if (!methods || methods.length === 0) return true;
return methods.some((m) => String(m).trim().toUpperCase() === method.toUpperCase());
}
function evaluate(rules, slug, category, verified, path, method) {
for (const r of rules || []) {
if (!matchApplies(r.match || {}, slug, category, verified)) continue;
if (!pathMatches(r.path_pattern || "", path)) continue;
if (!methodMatches(r.methods || [], method)) continue;
return r;
}
return null;
}
function verifyBundle(body, keys) {
let env;
try { env = JSON.parse(body); } catch { return null; }
if (!env || env.alg !== "ed25519") return null;
const pub = keys[env.kid];
if (!pub) return null;
const payload = Buffer.from(env.payload || "", "base64");
const sig = Buffer.from(env.sig || "", "base64");
if (sig.length !== 64) return null;
if (!crypto.verify(null, payload, pub, sig)) return null;
let bundle;
try { bundle = JSON.parse(payload.toString("utf8")); } catch { return null; }
if ((bundle.format_version || 1) > 1) return null;
return bundle;
}
function enforcer(cfg) {
const state = { bundle: null };
const counters = new Map();
async function refresh() {
try {
const keysRes = await fetch(cfg.keysUrl);
const keysJson = await keysRes.json();
const keys = {};
for (const k of keysJson.keys || []) {
const pk = toPublicKey(k.key);
if (pk) keys[k.kid] = pk;
}
if (Object.keys(keys).length === 0) return;
const res = await fetch(cfg.policyUrl, { headers: { "x-api-key": cfg.apiKey } });
if (res.status !== 200) return; // 304/other → keep last good bundle
const bundle = verifyBundle(await res.text(), keys);
if (bundle) state.bundle = bundle;
} catch {
/* fail-open: keep the last good bundle */
}
}
refresh();
setInterval(refresh, 5 * 60 * 1000).unref();
return function (req, res, next) {
try {
// Optional: exempt your own trusted traffic (e.g. logged-in users) from
// enforcement. Plug in your own check — it runs first, so a matching
// request is never blocked. Keep it cheap (inspect your session cookie /
// JWT), no I/O.
// if (isLoggedIn(req)) return next();
const b = state.bundle;
if (!b || b.mode !== "enforce") return next();
const ua = (req.headers["user-agent"] || "").toLowerCase();
let slug = "", category = "", best = 0, cidrs = [];
for (const a of b.agents || []) {
for (const tok of a.ua_tokens || []) {
const tl = String(tok).toLowerCase();
if (tl && ua.includes(tl) && tl.length > best) {
best = tl.length; slug = a.slug; category = a.category; cidrs = a.cidrs || [];
}
}
}
let verified = false;
if (slug) {
const ip = (req.socket && req.socket.remoteAddress) || "";
verified = cidrs.some((c) => ipInCidr(ip.replace(/^::ffff:/, ""), c));
}
const rule = evaluate(b.rules, slug, category, verified, req.path || req.url || "/", req.method);
if (!rule) return next();
if (rule.action === "deny") { res.statusCode = 403; reportBlock(cfg.apiKey, req, "deny"); return res.end("Forbidden"); }
if (rule.action === "throttle" && b.quota_enabled && rule.rate_limit) {
const key = slug || ("cat:" + category);
const { ok, retry } = throttleOk(counters, key, rule.rate_limit);
if (!ok) { res.statusCode = 429; res.setHeader("Retry-After", String(retry)); reportBlock(cfg.apiKey, req, "throttle"); return res.end("Too Many Requests"); }
}
return next();
} catch {
return next(); // fail-open
}
};
}
// reportBlock fire-and-forgets a report that this request was blocked, so the
// SilentShield dashboard's "blocked bots" report has data. Best-effort — errors
// are swallowed and it never delays the response.
function reportBlock(apiKey, req, outcome) {
try {
const s = {
ua: req.headers["user-agent"] || "",
ip: ((req.socket && req.socket.remoteAddress) || "").replace(/^::ffff:/, ""),
path: req.path || req.url || "/",
method: req.method,
outcome,
};
fetch("https://api.silentshield.io/api/v1/agent/telemetry", {
method: "POST",
headers: { "Content-Type": "application/json", "x-api-key": apiKey },
body: JSON.stringify({ sightings: [s] }),
}).catch(() => {}); // best-effort, fail-open
} catch {
/* ignore */
}
}
function throttleOk(counters, key, rl) {
const requests = rl.requests | 0, window = rl.window_sec | 0;
if (requests <= 0 || window <= 0) return { ok: true, retry: 0 };
const now = Math.floor(Date.now() / 1000);
const bucket = Math.floor(now / window);
let c = counters.get(key);
if (!c || c.bucket !== bucket) { c = { bucket, count: 0 }; counters.set(key, c); }
c.count++;
if (c.count > requests) {
const retry = Math.max(1, (bucket + 1) * window - now);
return { ok: false, retry };
}
return { ok: true, retry: 0 };
}
module.exports = { enforcer };
// Wiring (Express):
// const { enforcer } = require("./silentshield-enforcer");
// app.use(enforcer({
// policyUrl: "https://api.silentshield.io/api/v1/agent/policy",
// keysUrl: "https://api.silentshield.io/.well-known/silentshield-agent-keys",
// apiKey: "YOUR_API_KEY",
// }));WordPress
Using WordPress? The SilentShield plugin ships the enforcer from version 2.10.0 — no code needed. Turn it on under Advanced → "Block AI crawlers (enforce)" (off by default). It fetches and verifies the same signed policy server-side and blocks disallowed bots for you.
Next.js and other edge / serverless hosts
Next.js middleware runs on the Edge runtime, where verifying the signed bundle is impractical. For Next.js — and any edge or serverless host — put the reverse-proxy sidecar (below) in front of your app so enforcement happens before the request reaches it, or use the Go or Node.js enforcer above if you run your own server.
Reverse Proxy (nginx / Caddy / Traefik)
One tiny sidecar binary serves all three proxies: the proxy asks it once per request via forward auth. Start it with your site key and the trusted keys:
AGENT_POLICY_URL="https://api.silentshield.io/api/v1/agent/policy" \
AGENT_SITE_KEY="YOUR_API_KEY" \
AGENT_TRUSTED_KEYS='[{"kid":"…","key":"<base64 ed25519 pubkey>"}]' \
AGENT_SIDECAR_LISTEN=":8127" \
./agent-sidecarThe sidecar answers on GET /auth with 204 (allow), 403 (deny), or 429 plus Retry-After (throttle). Wire it into your proxy:
location = /_ss_auth {
internal;
proxy_pass http://127.0.0.1:8127/auth;
proxy_pass_request_body off;
proxy_set_header Content-Length "";
proxy_set_header X-Forwarded-Method $request_method;
proxy_set_header X-Forwarded-Uri $request_uri;
proxy_set_header X-Forwarded-Host $host;
proxy_set_header X-Forwarded-For $remote_addr;
proxy_set_header User-Agent $http_user_agent;
proxy_set_header Signature $http_signature;
proxy_set_header Signature-Input $http_signature_input;
proxy_set_header Signature-Agent $http_signature_agent;
}
location / {
auth_request /_ss_auth;
# ... your normal proxy_pass to the origin ...
}example.com {
forward_auth 127.0.0.1:8127 {
uri /auth
copy_headers User-Agent Signature Signature-Input Signature-Agent
# Caddy sends X-Forwarded-Method/-Uri/-Host automatically
}
reverse_proxy origin:8080
}http:
middlewares:
silentshield:
forwardAuth:
address: "http://agent-sidecar:8127/auth"
authRequestHeaders:
- "User-Agent"
- "Signature"
- "Signature-Input"
- "Signature-Agent"Cloudflare Worker
The Cloudflare Worker enforcer verifies the policy bundle with Web Crypto Ed25519 and blocks before requests ever reach your origin:
cd enforcers/cloudflare-worker
npx wrangler secret put AGENT_SITE_KEY
# set AGENT_POLICY_URL + AGENT_TRUSTED_KEYS in wrangler.toml [vars]
npx wrangler deployOfficial SDKs
Prefer a maintained package over the copy-paste snippet above? Install one of our official SDKs — the same enforcer plus the verify and observe helpers, kept up to date:
npm install @forge12interactive/silentshield-sdk-jsgo get github.com/forge12interactive/[email protected]composer config repositories.silentshield vcs https://github.com/forge12interactive/silentshield-sdk-php
composer require forge12interactive/silentshield-sdkSource, issues and the full README for each SDK are on GitHub: silentshield-sdk-js · silentshield-go · silentshield-sdk-php
Honest Limits
Agentic browsers on residential IP addresses (Comet, ChatGPT Atlas, and similar) are technically indistinguishable from human visitors. No product can reliably identify them — SilentShield therefore deliberately makes no enforcement promise for this category.
robots.txt and emitted content signals are advisory: well-behaved crawlers follow them, but nothing forces them to. Technical enforcement happens only in the enforcers on this page — and even they act conservatively, denying only positively identified known agents so that legitimate visitors are never blocked.
Configure Your Policy
Which agent may do what is configured per API key in the agent hub: API Keys → select a key → Agent. Presets cover the common cases; every change becomes a newly signed bundle that your enforcers pick up automatically.