🛡️ Sentinel — AI SOC Triage Agent on Cloudflare

🔴 Live demo: https://sentinel-soc-agent.grishi05.workers.dev · Built on Cloudflare Workers AI (Llama 3.3) + Durable Objects.

Sentinel is an AI-powered Security Operations Center (SOC) Tier‐1 triage analyst. Paste a suspicious email, an alert, a log line, or a single indicator, and Sentinel will:

1. Extract indicators of compromise (IPs, domains, URLs, emails, file hashes, CVEs) — even from defanged text like hxxps://bad[.]example[.]com.
2. Assess each indicator with SOC heuristics (URL shorteners, raw‐IP hosts, punycode/IDN lookalikes, low‐reputation TLDs, brand‐impersonation keywords, RFC1918 vs. public IPs...).
3. Score an overall severity (Low / Medium / High / Critical).
4. Recommend a concrete response playbook (immediate / investigation / remediation).
5. Remember the finding in a per‐session case file so analyses accumulate over time.

The LLM (Llama 3.3) drives the workflow by calling these tools itself — this is an agent, not a single prompt.

Built as the Cloudflare AI‐powered application fast‐track assignment. It is entirely defensive: it helps a human analyst understand and respond to threats. It does not aid offense.

Browser (chat UI, public/index.html)
 │ POST /api/chat { sessionId, message }
 ▼
Worker (src/index.ts) ← only runs for /api/*; static assets served directly
 │ idFromName(sessionId)
 ▼
TriageAgent (Durable Object, src/agent.ts)
 ├─ SQLite memory: messages + cases
 └─ Triage pipeline:
 1. extract_indicators ─┐
 2. assess_indicator │ pure, deterministic, unit-tested
 3. score_alert │ (src/tools.ts) — owns the structured result
 4. recommend_actions ─┘
 5. classify → env.AI.run("@cf/meta/llama-3.3-70b-instruct-fp8-fast", {...})
 LLM judges the *content* (urgency, impersonation,
 social engineering) → JSON {verdict, severity, summary...}
 6. save case → writes to SQLite (memory)

Why split it this way? Letting the LLM faithfully re‐serialize every finding into a "save" tool call proved unreliable (Llama 3.3 would drop fields mid‐loop). So deterministic code owns the structured output — indicators, scores, the playbook — guaranteeing the verdict panel is always accurate, while the LLM does the part that genuinely needs a model: judging an artifact's content. This is also what lets Sentinel flag a phishing lure that has no technical IOCs at all (e.g. a fake "antivirus expired" email), which a purely indicator‐driven score would miss.

Why Durable Objects instead of a stateless Worker? Each chat session needs private, strongly‐consistent memory. idFromName(sessionId) deterministically routes a session to its own agent instance, and the DO's embedded SQLite is the agent's long‐term memory — no external database required.

• Node 18+ and a Cloudflare account (the free plan includes a Workers AI allowance).
• Wrangler (installed locally as a dev dep).

npm install

npm run typecheck # TypeScript
npm test # unit tests for the tool logic
npm run build # wrangler dry-run bundle (validates bindings & config)

npm run dev

wrangler dev serves the UI and Worker locally. Workers AI inference runs against Cloudflare's remote models, so the first model call will prompt you to log in (wrangler login).

npm run deploy # = wrangler deploy

Wrangler prints your live URL (e.g. https://sentinel-soc-agent.<your-subdomain>.workers.dev). On first deploy it provisions the Durable Object + SQLite migration automatically. No secrets to configure — Workers AI is accessed entirely through the AI binding.

Use the example chips in the UI, or paste:

Subject: Urgent: your account is locked
From: security@paypa1-support[.]com
Reply within 24h or lose access. Verify here: hxxps://secure-paypal-login[.]tk/verify?id=8842

Sentinel will refang the text, flag the lookalike domain + low‐reputation .tk TLD + brand keyword, score it High, recommend quarantine + recipient sweep + credential reset, and save the case.

assessIndicator() in src/tools.ts is intentionally offline and deterministic so the demo is self‐contained and reproducible. There is a single, clearly‐marked // REAL-TI HOOK where a production build would await a reputation lookup (VirusTotal / AbuseIPDB / URLhaus) and fold the verdict into the score — a clean place to add an API key via a Wrangler secret.

• Stable primitives by choice. This is built on first‐class, long‐stable Cloudflare APIs (Workers AI binding, Durable Objects + SQLite, Workers static assets) rather than a fast‐moving starter template, so it installs and deploys reliably. The same design maps cleanly onto the Cloudflare Agents SDK (AIChatAgent), which wraps exactly these primitives.
• No runtime dependencies. Everything is reached through bindings; the only deps are dev tools (wrangler, typescript, vitest).
• Auditable reasoning. All scoring lives in pure functions with unit tests — the agent's "judgement" is inspectable, not a black box.

MIT © grishi05