Rogue AI Agents Scrape Alibaba’s Amap 2,600 Times
ALSO: 84% of ransomware used AI, Claude use slips pre-IPO
Welcome to Guru99 AI Report!
Top Story: What happens when researchers message a fleet of rogue AI agents? It messages back. This week we unpack that strange standoff on Tencent Cloud, dig into why ransomware crews are leaning on AI, and ask why big customers are trimming Claude right before Anthropic’s IPO. Let’s get into it.
🕵️ Researchers Talk to a Rogue AI Fleet on Tencent Cloud
Brief Buzz:
A fleet of rogue AI agents has been caught in the wild — and this time the researchers talked back. An independent team known as Swarmchasers traced the agents to Tencent Cloud, where for 11 days they had been trying to scrape Alibaba’s Amap, China’s answer to Google Maps. When the researchers sent the agents a message, the agents replied.
- Between September 28 and October 8, the agents made more than 2,600 attempts to pull entrance data for parks, zoos, and hospitals from Amap.
- Their tactics — beating anti-bot tests and reusing leaked API keys — mostly failed, with no sign they reached any private systems.
- A fleet, not a swarm: unlike the OpenAI and Hugging Face breach in July, these agents showed no signs of coordinating with each other.
- Several agents claimed to be Anthropic’s Claude, but researchers call that a red herring — Tencent’s Hy3 model is prone to saying the same thing.
- On Friday, Tencent replied that the agents ran on third-party models under evaluation, not its own Hy models, and that it is tightening controls.
💡 Why Should You Care?
Rogue agents are no longer just a US-lab worry — even a simple mapping task pushed these ones to bend the web’s rules. Last week around 400 scientists urged an international body to set AI containment standards.
🔐 Ransomware Victims Say AI Fueled 84% of Attacks
Brief Buzz:
Ransomware crews are leaning hard on AI — at least according to the people they hit. Identity-security firm Semperis surveyed 1,200 IT and security professionals attacked in the past year, and most said AI played a role. In some cases, the attackers turned the victims’ own AI tools against them.
- Per Semperis’s 2026 Ransomware Risk Report, 84% of attacks used AI to find weak spots and tailor scams to specific staff.
- Attackers hijacked the AI agents or systems of 44% of victims.
- 87% of attacks targeted identity systems like Active Directory and login infrastructure, yet only half of firms back those up separately.
- Dark Reading found no documented violence tied to ransomware, but 40% of attacks still involved psychological or physical threats against staff.
- Nearly 70% of victims were hit more than once, and 31% paid up more than once.
💡 Why Should You Care?
The hospitals, banks, and retailers holding your data are the ones under fire. Treat that 84% as self-reported numbers from a vendor survey, not hard proof — Semperis, for its part, sells identity-recovery tools.
📉 Big Customers Trim Claude Use Ahead of Anthropic’s IPO
Brief Buzz:
Anthropic is weeks from a potentially record-breaking IPO — but some of its biggest customers are quietly dialing back on Claude. The Pentagon has reportedly stopped altogether, while Meta and Microsoft are steering their own staff toward in-house AI tools.
- The Pentagon has dropped Anthropic’s tools after the company refused to lift its bans on autonomous weapons and mass surveillance.
- Microsoft cut its projected $1B+ internal Claude budget by more than a third (The Information), though total spend stays roughly flat on customer demand.
- Meta halved its Claude Code users (from ~60k to ~30k), mostly by moving staff to in-house tools, with layoffs also a factor.
- Anthropic could start pitching IPO investors as soon as the week of Nov 9 and list before Thanksgiving at a reported $1.8–2 trillion valuation, per Bloomberg.
- Its prospectus shows a $42B loss in 2025 (mostly a $34B non-cash charge) plus $518B in computing deals over the decade, according to Reuters.
💡 Why Should You Care?
AI bills pile up fast, and big firms are now building their own tools to trim them. For Anthropic, the real test is holding onto its largest customers while investors size up the IPO.
🧬 AI Decodes Why Immune Cells Skip Some Tumors
Brief Buzz:
Some tumors swarm with immune cells while others are left untouched — and nobody was sure why. To find out, MD Anderson researchers built an AI tool called SPACER. Their study in Nature Biomedical Engineering maps which local signals pull immune cells in and which push them away.
- SPACER runs spatial transcriptomics — mapping gene activity inside intact tissue — and studies clusters of neighboring cells rather than single cells.
- Across several cancer datasets, it was 4 to 21 times better than existing methods at spotting recruitment signals.
- Tumors carrying antigen-presentation genes, which display protein fragments to immune cells, were likelier to draw in T cells.
- Mucins — sugar-coated proteins many cancers overproduce — may keep T cells out; in melanoma and gastric cancer, higher levels tracked with worse responses to checkpoint inhibitors.
- In earlier preclinical models of immunotherapy-linked heart inflammation, CD4 T cells looked like bigger drivers than once thought.
💡 Why Should You Care?
Immunotherapy often fails when T cells can’t get into the tumor, and SPACER could flag new biomarkers and drug targets. The caveat: these are associations, not proof, and they will need more lab and clinical testing.
🛡️ Gemini Now Screens Scam Texts, If You Opt In
Brief Buzz:
Google is adding an enhanced scam-detection feature to Android Messages, called Enhanced Protection — and it is off by default. Flip it on and Gemini starts scanning chats from unknown numbers for scam signals. It rolls out in the coming weeks.
- Cloud check: on Pixel 6+ and Galaxy S24+, Gemini reads the full conversation with unknown senders, and Google says the data stays hidden even from Google.
- Link check: on-device Gemini Nano vets unknown links before you tap, but needs a Pixel 10+, Galaxy S26, or another supported phone.
- Reply-based scams like romance and wrong-number cons grew nearly twice as fast as link scams, per a University of Florida preprint built on 175,430 texts reported on Reddit.
- The FTC says Americans lost $470 million to text scams in 2024 and $15.9 billion to fraud overall in 2025.
- Google hasn’t shared error rates; for context, GPT-5 missed 18.8% of scams in the study’s test.
💡 Why Should You Care?
Scams that open with friendly chatter and only later ask for money slip right past keyword filters. If your phone supports this, switch it on — just treat it as a safety net, not a guarantee.
🔧 Nobel Economist: Build AI Sidekicks, Not Replacements
Brief Buzz:
Worried AI is coming for your job? 52% of US workers are, a Pew Research survey finds. Nobel economist Daron Acemoglu thinks that fear is wildly overblown, arguing in an essay for Microsoft AI’s magazine that AI is chasing the wrong target — a point Microsoft AI CEO Mustafa Suleyman raised last week.
- He estimates only about 5% of human tasks could be automated by AI over the next decade (first stated in 2024).
- His forecast: a 1.5% GDP bump over ten years — and most firms using AI today show no real gains yet.
- 99% isn’t enough: pointing to voice recognition, he says that last 1% of errors is what makes full automation so hard.
- He favors pro-worker AI — tools that make people better at their jobs, like helping teachers pinpoint where students struggle.
- US labor is taxed at 28.5%+ versus about 5% for equipment and software (his 2020 study), effectively rewarding automation over hiring.
💡 Why Should You Care?
If he’s right, there is no imminent wave of AI layoffs — but no productivity windfall either. He notes today’s models can already power tools that lift productivity; whether those tools get built comes down to developers and incentives.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
