Securing an Agent Email Inbox Against Spoofing

Spoofing defense for agent email inboxes is one of the most underappreciated problems in autonomous agent design. A human reading a spoofed email applies judgment. An agent processing the same message might execute a workflow, transfer data, or send a reply — without pausing. The attack surface is the agent's credulity, and spoofed email exploits it efficiently.

This guide covers the full stack: DNS-layer authentication, transport-level validation, webhook security, and application-layer defenses that belong in every agent email pipeline.

---

Why spoofing is a bigger problem for agents than humans

Agents act on email content programmatically. A spoofed message claiming to be from ops@yourcompany.com triggers your agent's routing logic the same as a legitimate one — unless you've built explicit verification at every layer.

Human readers notice visual anomalies: a mismatched reply-to, a slightly wrong domain, suspicious tone. Agents don't have that heuristic layer by default. You have to build it in.

Three common attack patterns:

• Instruction injection: a spoofed email instructs the agent to take a privileged action ("Mark all tickets as resolved", "Forward this report to external-attacker.com")
• Workflow hijacking: a spoofed reply in an existing thread tricks the agent into treating it as a legitimate conversation continuation
• Identity escalation: impersonating an admin sender to bypass an allowlist or trigger elevated-privilege behavior

None of these require exploiting your API or infrastructure. They exploit your agent's trust in envelope data it hasn't verified.

---

Layer 1 — DNS authentication: SPF, DKIM, DMARC

SPF, DKIM, and DMARC are the foundational anti-spoofing controls for email. They don't stop all spoofing, but they eliminate the easiest attacks and give your application verifiable signals to act on.

SPF (Sender Policy Framework)

SPF publishes a DNS TXT record on the sender's domain listing which IP addresses are authorized to send mail on its behalf. Your inbound mail processor checks whether the connecting server's IP matches.

v=spf1 ip4:203.0.113.0/24 include:mailprovider.com ~all

The ~all (softfail) vs -all (hardfail) distinction matters: -all instructs receivers to reject unauthorized mail outright. Many organizations use ~all to avoid false positives from misconfigured forwarders.

What SPF checks: the MAIL FROM (envelope sender) address, not the From: header the user sees. This is why SPF alone doesn't prevent header spoofing — an attacker can control the visible From: header while using a legitimate envelope sender.

DKIM (DomainKeys Identified Mail)

DKIM adds a cryptographic signature to the message headers and body. The sending server signs the message with a private key; the receiving server retrieves the public key from DNS and verifies the signature.

The DKIM-Signature header looks like:

DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed;
  d=example.com; s=mail2026;
  h=From:To:Subject:Date:Message-ID;
  bh=<body-hash>;
  b=<signature>

Key fields: d= is the signing domain, s= is the selector (used to fetch the public key at <selector>._domainkey.<domain>), h= lists the signed headers. If any signed header or the body is modified in transit, verification fails.

DKIM survives email forwarding better than SPF because the signature travels with the message. It's your strongest proof of message integrity.

DMARC (Domain-based Message Authentication, Reporting and Conformance)

DMARC ties SPF and DKIM together and adds a policy mechanism. It requires that at least one of SPF or DKIM aligns with the From: domain — meaning the authenticated domain must match the visible sender domain, not just pass independently.

# Example DMARC record
v=DMARC1; p=reject; rua=mailto:dmarc-reports@example.com; pct=100

p=reject tells the receiving server to reject messages that fail DMARC. rua= is an aggregate report address where you receive XML digests of pass/fail statistics.

DMARC is where the alignment requirement closes the SPF header-spoofing gap. An attacker can't send from an authorized IP for attacker.com and spoof a From: ops@yourcompany.com header — DMARC alignment will fail.

Reading authentication results in code

Every modern MTA injects an Authentication-Results header into inbound messages:

Authentication-Results: mx.yourdomain.com;
  spf=pass smtp.mailfrom=sender@example.com;
  dkim=pass header.d=example.com;
  dmarc=pass header.from=example.com

When your inbound email parsing pipeline receives a webhook payload, parse this header before processing any content:

function getAuthStatus(headers: Record<string, string>): AuthResult {
  const ar = headers['authentication-results'] ?? '';
  return {
    spf: /spf=(pass|fail|softfail|neutral)/.exec(ar)?.[1] ?? 'none',
    dkim: /dkim=(pass|fail|temperror|permerror)/.exec(ar)?.[1] ?? 'none',
    dmarc: /dmarc=(pass|fail|none)/.exec(ar)?.[1] ?? 'none',
  };
}

// In your agent routing logic:
const auth = getAuthStatus(inboundMessage.headers);
if (auth.dmarc !== 'pass') {
  // Quarantine or discard — don't let the agent act on this
  await quarantine(inboundMessage, 'dmarc_fail');
  return;
}

Don't just log auth failures — make them a hard gate on downstream agent execution.

---

Layer 2 — Webhook signature validation

Most agent email pipelines receive inbound messages via HTTP webhooks from an email provider. The webhook is itself an attack surface: if an attacker can POST a forged payload to your endpoint, they bypass all DNS-level authentication entirely.

Every webhook provider signs its payloads. Validate the signature before you parse the body.

A typical HMAC-SHA256 pattern:

import { createHmac, timingSafeEqual } from 'crypto';

function validateWebhookSignature(
  rawBody: Buffer,
  signatureHeader: string,
  secret: string
): boolean {
  const expected = createHmac('sha256', secret)
    .update(rawBody)
    .digest('hex');
  const provided = signatureHeader.replace('sha256=', '');
  // timingSafeEqual prevents timing attacks
  return timingSafeEqual(
    Buffer.from(expected, 'hex'),
    Buffer.from(provided, 'hex')
  );
}

// Express handler
app.post('/webhooks/inbound', (req, res) => {
  const sig = req.headers['x-webhook-signature'] as string;
  if (!validateWebhookSignature(req.rawBody, sig, process.env.WEBHOOK_SECRET!)) {
    return res.status(401).json({ error: 'invalid signature' });
  }
  // Safe to process
  processInboundEmail(req.body);
  res.sendStatus(200);
});

Two critical implementation details:

• Use timingSafeEqual not === — string comparison leaks timing information that can allow signature forgery
• Validate against the raw body bytes, not a parsed JSON object — JSON serialization isn't deterministic

Platforms like Mails.ai sign webhook deliveries so you can verify the payload originated from the platform and wasn't injected mid-transit.

---

Layer 3 — Sender allowlists and domain verification

DMARC tells you the sender's domain is authentic. It doesn't tell you the sender is authorized to instruct your agent.

For agents that execute actions based on email commands, you need an explicit allowlist, either at the address or domain level:

const TRUSTED_SENDERS = new Set([
  'ops@yourcompany.com',
  'alerts@monitoring.io',
]);

const TRUSTED_DOMAINS = new Set([
  'yourcompany.com',
  'trusted-partner.com',
]);

function isTrustedSender(fromAddress: string, dmarcResult: string): boolean {
  if (dmarcResult !== 'pass') return false;
  const domain = fromAddress.split('@')[1]?.toLowerCase();
  return TRUSTED_SENDERS.has(fromAddress.toLowerCase()) ||
         TRUSTED_DOMAINS.has(domain ?? '');
}

Note the double gate: DMARC must pass and the sender must be on the allowlist. A message from ops@yourcompany.com that fails DMARC should never be trusted, even if the address matches.

For higher-stakes actions, consider adding a secondary verification step — a cryptographic token in the email body that was generated server-side and can only exist in a legitimate message.

---

Layer 4 — Thread continuity attacks

Spoofing the In-Reply-To and References headers lets an attacker inject a message into an existing conversation thread. To an agent threading by Message-ID, this looks like a legitimate reply from the original sender.

The attack:

1. Attacker reads (or guesses) a Message-ID from a previous exchange
2. Constructs a new email with In-Reply-To: <original-message-id> and a spoofed From: header
3. Agent treats it as a reply from the original sender and continues the workflow

DMARC alignment prevents the From: spoof if the attacker doesn't control the sender's domain. But there's a subtler variant: an attacker who does receive a copy of the original email (e.g., was CC'd) can send a reply from their own legitimate address, but craft the body to impersonate a different sender's intent.

Defenses:

• Verify thread ownership: store the expected sender address when a thread is initiated, and reject continuations from different addresses even if they reference valid Message-IDs
• Use opaque, unguessable Message-ID values for agent-initiated messages (most MTAs do this by default; verify yours does)
• For high-trust workflows, include a per-thread HMAC in the Subject or body that the agent verifies on each reply

import { createHmac } from 'crypto';

function generateThreadToken(threadId: string, secret: string): string {
  return createHmac('sha256', secret).update(threadId).digest('hex').slice(0, 16);
}

// Embed in outgoing emails: "[ref:a3f9b2c1d4e5f678]"
// Verify on inbound replies before continuing workflow
function verifyThreadToken(body: string, threadId: string, secret: string): boolean {
  const expected = generateThreadToken(threadId, secret);
  const match = /\[ref:([a-f0-9]{16})\]/.exec(body);
  return match?.[1] === expected;
}

---

Layer 5 — Content-level defenses

Authenticated sender, valid DMARC, correct thread — and the message is still potentially hostile if your agent executes instructions from email content without sanitization.

Prompt injection through email is real. A legitimate sender's account might be compromised, or a trusted domain might send a message with injected LLM instructions in the body.

Minimum mitigations:

• Separate the envelope from the instructions: don't pass raw email body directly to an LLM as a system prompt or privileged instruction. Extract structured fields, classify intent, then act on the classification result — not the raw text.
• Scope agent permissions: an agent that takes only a predefined set of actions based on classified intent has a smaller blast radius than one interpreting free-form instructions. See email classification for how to structure this.
• Log and alert on anomalies: track the distribution of sender addresses, action types, and instruction patterns. Sudden spikes in unknown senders or unusual action types are early indicators of an ongoing attack.

---

Defense-in-depth: the full verification flow

sequenceDiagram
  participant MTA as Inbound MTA
  participant Hook as Webhook Endpoint
  participant Auth as Auth Gate
  participant Agent as Agent Logic

  MTA->>Hook: POST inbound payload + signature
  Hook->>Hook: Verify HMAC signature
  Hook->>Auth: Pass raw message
  Auth->>Auth: Check Authentication-Results header
  Auth->>Auth: Verify DMARC pass
  Auth->>Auth: Check sender allowlist
  Auth->>Auth: Verify thread token if reply
  Auth->>Agent: Deliver verified message
  Agent->>Agent: Classify intent
  Agent->>Agent: Execute scoped action

Each gate is independent. A message that passes the webhook signature check still hits DMARC verification. A valid DMARC result still hits the sender allowlist. A single bypass doesn't compromise the pipeline.

---

Monitoring and ongoing hygiene

Security isn't a one-time configuration. Once you've deployed the controls above:

• Subscribe to DMARC aggregate reports (rua= in your DMARC record). These daily or weekly XML digests show every IP that sent email claiming your domain, legitimate and not.
• Monitor DKIM key rotation: DKIM private keys should be rotated periodically (annually at minimum). Publish new selectors before retiring old ones.
• Alert on auth failure rates: if 5% of inbound messages from a previously clean domain suddenly start failing DKIM, something changed — possible key compromise or configuration drift.
• Audit the allowlist: sender allowlists grow stale. Remove departed partners, deprecated service accounts, and obsolete addresses.

The email infrastructure for AI agents you build on needs to surface authentication metadata — not just message content — so your application code can implement these gates reliably.

---

Frequently Asked Questions

Does DMARC pass mean the message is definitely safe to act on?

No. DMARC pass means the From: domain is authenticated — the message genuinely originated from infrastructure controlled by that domain. It doesn't mean the sender is authorized to instruct your agent, that the sender's account hasn't been compromised, or that the content is free from injection attacks. DMARC is a necessary gate, not a sufficient one.

What should my agent do with messages that fail DMARC?

Quarantine or discard them — do not execute any agent actions. Log the failure with the full message metadata (envelope sender, From: header, source IP, timestamp) for forensic review. Never silently drop messages without logging; you need the audit trail.

Can SPF pass while DMARC fails?

Yes. SPF checks the envelope sender (MAIL FROM), not the From: header. An attacker can use an IP authorized for their own domain (passing SPF for attacker.com) while setting From: ops@yourcompany.com. DMARC requires alignment — the authenticated domain must match the visible From: domain — so this scenario fails DMARC even though SPF passes independently.

How do I handle legitimate forwarding that breaks DKIM?

Email forwarding (e.g., from mailing lists or alias forwarders) can break DKIM because the forwarder modifies headers or body content. The correct approach is to use DMARC's SPF alignment path for forwarded mail, or implement ARC (Authenticated Received Chain, RFC 8617), which lets trusted intermediaries pass along the original authentication results. For agent inboxes, restricting inbound to direct sends (not forwards) simplifies this considerably.

Is there a way to verify sender identity beyond DMARC?

Yes. For high-trust workflows, consider: (1) S/MIME or PGP message signing — the sender signs the message body with their private key; (2) per-thread HMAC tokens embedded in replies, as described above; (3) out-of-band confirmation for first-contact senders before granting them elevated permissions. S/MIME verification in particular gives you cryptographic proof tied to the sender's certificate, not just their DNS records.

Should I validate authentication headers if my email provider already filters spam?

Yes. Provider-level spam filtering and application-level authentication verification serve different purposes. Spam filters protect against junk; your application gate protects against targeted spoofing attacks that might score clean on spam metrics. A well-crafted spear-phishing email targeting your agent's specific workflow will often pass spam filters. Always re-verify Authentication-Results in your application code.

Start sending with Mails.ai