On September 11, Tedium published a piece on Kaixin Tang, founder of iLands. He'd been using an AI Agent to auto-generate and blast out highly personalized sales emails, targeting tens of thousands, maybe hundreds of thousands, of prospects.
The story hit the Hacker News front page that day, scoring 43 points and 19 comments. The comment section split into two camps: "Isn't this just spam?" and "Okay, technically impressive."
After reading the source, what struck me was different: this pipeline has an absurdly low barrier to entry. Anyone who can write Python could rebuild it over a weekend.
What iLands actually does: dressing mass outreach as 1-on-1
Traditional bulk email looks like this:
Dear Valued Customer, our product can help your business...
Instant spam filter.
What iLands's Agent produces looks like this:
I read your post from last year on Postgres index optimization, especially the section on BRIN. We recently applied a similar approach for an e-commerce client and saw a 4x improvement in write throughput. If you're interested, I can send over the detailed case study.
Most recipients' first reaction: "This person actually read my blog."
But here's what actually happened: three seconds before sending, the Agent scraped the target's personal site, GitHub, LinkedIn, and Twitter, stuffed all the public info into a prompt, and let the LLM generate a reply that reads like it was written by hand. Zero human involvement.
That's the core of the Tedium story: the Agent pushed spam to near-human quality.
The Agent pipeline, broken down: three fully automated stages
Split open the outreach flow, and it comes down to three stages: research → draft → send. Each stage maps to a specific Agent capability.
Stage 1: Research (target profiling)
Given a domain or LinkedIn URL, the Agent automatically:
- Scrapes personal sites, blogs, GitHub profiles
- Extracts titles and summaries from the most recent 3–5 posts
- Pulls LinkedIn work history and education background
- Pulls recent Twitter/X posts
The stack is roughly Playwright or headless Chrome for crawling, plus a simple LLM call for information extraction. Here's a simplified version:
from playwright.sync_api import sync_playwright
import anthropic
def scrape_target(url: str) -> dict:
with sync_playwright() as p:
browser = p.chromium.launch(headless=True)
page = browser.new_page()
page.goto(url, wait_until="networkidle")
html = page.content()
browser.close()
client = anthropic.Anthropic()
msg = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"从以下 HTML 提取这个人的职业背景、最近写的主题、技术栈偏好:\n\n{html[:50000]}"
}]
)
return msg.content[0].text
profile = scrape_target("https://example.com/blog")
That's the whole thing. The crawl-plus-extraction pipeline runs under 80 lines.
Stage 2: Draft (copy generation)
Feed the profile from stage one into the LLM. The prompt looks something like this:
你是一名 SaaS 销售,要给下面这个人写一封冷邮件:
姓名:{name}
职业背景:{profile['background']}
最近关注的话题:{profile['topics']}
要求:
- 200 字以内
- 引用他最近一篇文章或一个观点(具体到细节)
- 不要使用"尊敬的"、"我们公司"这种套路词
- 口吻像一个同行在交流,不是销售在推销
- 结尾留一个低门槛的 CTA(比如"如果你有兴趣可以看看这个 case")
Claude Sonnet 4.5 or GPT-5 handles this for roughly 2–3 cents per email.
Stage 3: Send (delivery)
The least technical stage, but also the easiest place to blow up:
- Rotate across multiple SMTP accounts (so a single IP doesn't get flagged)
- Randomize send timing (30–180 seconds between emails)
- Configure DKIM / SPF / DMARC properly (most cheap ESPs handle this)
- Spread across domains (10 domains × 50 mailboxes = 500 sending identities)
Why Gmail's spam filter doesn't catch it
Gmail's spam model combines content signals and behavioral signals.
Content dimension: Classic spam is templated language, hyperbolic promises, link farms. Every email the iLands Agent generates is unique, so the probability of hitting spam keyword rules is near zero.
Behavioral dimension: Traditional spam is a single wave: one IP, one domain, thousands of emails sent simultaneously. The Agent pipeline does three things to get around this:
- Domain pool: The report mentions Tang rotates through a set of domains, each one sending only 20–50 emails per day.
- Human-like pacing: Randomized intervals, sending within business hours, simulating the 10am–4pm peak pattern.
- Reply-rate optimization: Mailboxes that get opened or replied to get flagged, and unopened recipients see reduced frequency going forward.
Combined, what Gmail sees looks like "a normal salesperson doing regular email outreach."
It's not bulletproof. HN commenters pointed out: a WHOIS lookup on the sender's domain reveals registration under six months old, no website, email-only purpose — classic spam infrastructure. But the average recipient won't bother checking.
The same Agent abuse is already spreading
iLands isn't an isolated case. The HN thread surfaced a bunch of similar patterns:
- SEO backlink Agents: auto-crawl blog lists, AI-generate comments, drop links in bulk
- LinkedIn connection Agents: use Browser Use to drive headless Chrome, simulate real clicks, send 200 requests per day
- Comment-spam Agents: auto-reply on Hacker News, Reddit, and Product Hunt with product links attached
- Job-description Agents: HR teams blasting templated "we reviewed your resume" messages to candidates
Every scenario follows the same technical chain: crawl → LLM rewrite → human-mimicking delivery.
Why the platforms aren't moving
Why doesn't Gmail / Outlook / LinkedIn just block this? Three reasons, each uglier than the last.
1. False-positive cost. Legitimate salespeople really do send bulk emails — they just don't have AI helping them write. If the filter gets too aggressive, legal B2B sales dies too. Google won't take that bet.
2. Detection difficulty. Agent-generated emails and human-written emails are now indistinguishable at the text level. To tell them apart, you'd have to analyze the behavioral chain — which means Gmail monitoring your browser activity, scroll patterns, mouse movements. A different kind of disaster.
3. Commercial incentive. Gmail makes money on enterprise plans, and B2B sales email open rates are a KPI. LinkedIn Premium sells unlimited InMail. The platforms profit from spam.
My read: waiting for platform governance is a mistake. They have neither the motivation nor the capability — or at least, they're unwilling to build that capability through surveillance.
Editorial take
The interesting part of the iLands story isn't "yet another spam incident."
What it proves is this: Agents have pulled individual bad actors up to organizational-scale output. Pulling off large-scale personalized spam used to require hiring ten salespeople, each writing 50 emails a day. Now one person plus an Agent pipeline outproduces all ten.
The more important point: the marginal cost is close to zero. One more target address costs a few cents in LLM calls. There's no ceiling on scale.
My bet for the next twelve months: dedicated anti-Agent detection services will emerge. Likely forms:
- Browser plugins: analyze the email you're currently reading in real time, flag the "AI-generated probability"
- Enterprise gateways: run reverse-LLM detection and sender behavioral profiling on inbound mail
- Reverse-crawling services: look up the sender's domain / email WHOIS, registration date, linked accounts
The first wave will probably lean on text-statistics approaches like GLTR or DetectGPT. But it'll quickly move up to behavioral chain analysis — checking whether the sender is using rhythmic sending patterns, whether the domain is freshly registered, whether the IP range matches common email-provider egress blocks.
The open question: how will false-positive rates evolve? If these tools flag 1% of legitimate B2B email, enterprises won't adopt them. Get it under 0.1% and it could become standard SaaS infrastructure.
Another open question: what's the legal definition? GDPR is strict about "automated decision-making," but the US has no equivalent federal law. At the state level, California's CCPA grants the right to "opt out of automated processing," but enforcement is essentially zero. In the short term, the law won't catch up with the tech.
What's different this time: this spam isn't "high-volume" — it's "looks human-written." Every legacy anti-spam system built its defenses around volume and template signatures. Agents bypassed both. This is a rule-level failure, not a technical-detail failure.
Sources:
- Tedium: iLands and the rise of Agent-driven personalized email outreach
- Hacker News discussion (item 49671159)