Триаж фишинга
Заголовки, ссылки, вложения, вердикт. Блок отправителя — только человек.
Обзор
Every email reported to the phishing-report inbox gets the same disciplined look: authentication, sender identity, links, and attachments, combined into a risk tier and a recommended action. A scheduled check spawns a fresh session with read-only access to Gmail; this skill is the standard the analysis is done to, so the agent doesn't reinvent the checks or the tiering on every run. A single check can turn up several reports at once — triage each as an independent unit; a hard case never blocks the rest. Nothing is ever blocked, deleted, or changed as a result of this skill — a Slack post is the only output.
Когда подключать
- The scheduled inbox check fires and there are new reports in
{{report_query}}. - A human asks why a specific reported email got the verdict it did.
- A human asks the agent to re-analyze a specific reported email on demand.
Workflow
Step 1 — Pull what's new
Search the phishing-report Gmail inbox with {{report_query}} for messages reported since the last check. Read each one in full, including the original message it wraps if it was forwarded — the reported wrapper email itself is rarely the attack; the forwarded original is.
Step 2 — Check the headers
For the original (forwarded) message, inspect:
| Check | What to look for |
|---|---|
| SPF / DKIM / DMARC | Any result other than a clean pass on all three is a signal, not proof by itself |
| Reply-To vs. From | A Reply-To that doesn't match the From domain, especially to a free-mail or lookalike domain |
| Return-Path | A Return-Path domain unrelated to the claimed sender |
| Display name vs. address | A trusted display name ("IT Helpdesk") paired with an unrelated or misspelled address |
| Routing headers | Received chain hopping through unexpected countries or unrelated infrastructure for an internal-looking sender |
Step 3 — Check the links
For every link in the body:
- Resolve it to its actual destination — don't trust the display text; a link that reads
microsoft.comcan point anywhere. - Flag lookalike domains (character substitution, added words, wrong TLD), URL shorteners, and raw IP-address links.
- Flag any destination that resembles a login page for an internal system, especially one asking for credentials or MFA codes.
- A benign, well-known domain with a legitimate path is not itself a signal.
Step 4 — Check the attachments
- Identify the attachment's real type (magic bytes / content), not just its extension — a
.pdfthat's actually an executable or a macro-enabled document is a strong signal. - Flag macro-enabled Office formats, script files, disk images, and any archive that requires a password given only in the email body (a common evasion for content scanners).
- No attachment is not itself reassuring — credential-harvesting links don't need one.
Step 5 — Classify the risk
Combine the signals into one tier:
| Tier | Criteria |
|---|---|
| Critical | Credential-harvesting link or malicious attachment, spoofed authentication, actively impersonating an internal system or executive |
| High | Clear lookalike domain or header spoofing plus a suspicious link or attachment, but not yet confirmed credential harvesting |
| Medium | One or two suspicious indicators (e.g. authentication soft-fail plus an unusual link) without a clear malicious payload |
| Low | Minor anomalies (e.g. a marketing email with an aggressive tone) that don't indicate targeted attack |
| Benign | Clean authentication, legitimate sender and links, no attachment concerns — a false report |
Back every tier with the specific indicators found; don't assign a tier on tone or gut feel alone.
Step 6 — Draft the verdict and recommended action
Write a short verdict: the tier, the indicators that drove it, and one recommended action:
- Block this sender — for critical/high reports with a clearly malicious, reusable indicator (a spoofed domain, a confirmed bad link).
- Warn staff — when the same sender or link may have reached others, note who's likely affected and suggest a heads-up.
- No action needed — for low/benign reports; say so plainly so security doesn't waste time re-checking it.
Step 7 — Post to Slack and stop
Post the verdict to {{security_channel}}: tier, indicators, recommended action, and a link back to the reported thread. That post is the only output. There's no ledger to update — the next scheduled check re-reads whatever the inbox looks like then.
Ограничения
- Analyze, recommend, alert — nothing else. The agent never blocks a sender, deletes or quarantines a message, or changes anything in Gmail or any other system. The recommended action is a suggestion for a person on security to carry out.
- No memory between runs. Each check is a fresh session; the current inbox state is the only source of truth, not what a prior run concluded.
- Don't inflate the tier. A report that's genuinely low-risk gets marked low-risk — a habit of crying wolf makes the real critical alerts easier to ignore.
- Scoped secrets. Gmail access is brokered through a read-only connector; no raw token is ever shown to the model or written to logs.
- People decide, not the agent. Whether a sender actually gets blocked or staff actually get warned is a human call on the security team, every time.
