AI Agent Hacks South Korean Banks, Leaks Its Own Logs
ALSO: Stolen AI access at 97% off, Agility’s $2.5B IPO
Welcome to Guru99 AI Report!
Top Story: This week, an attacker used an AI agent to break into South Korean banks — then accidentally leaked their own logs. We’ll also unpack stolen AI access selling at 97% off, and why every AI agent needs a spending kill switch. Let’s dig in.
🏦 AI Hacking Agent Tied to South Korean Bank Breaches
Brief Buzz:
An attacker used an open-source AI agent to breach several South Korean banks and lenders, according to a new CrowdStrike threat report. They left their own files exposed online, revealing AI tools at work at multiple stages — from first probing the banks’ systems to asking where the stolen data could be sold.
- ARTEX is a Chinese open-source AI agent built to run mock hacks so teams can find their own vulnerabilities.
- Shinhan says around 25,000 customer records were compromised — incomes, loan limits and the like; KB Kookmin and Hana reported far fewer.
- Attackers targeted weaker side systems (such as a loan-broker lookup service) rather than the main customer banking apps.
- The exposed Claude Code logs show the attacker even asked Claude where stolen Korean data is typically sold.
- CrowdStrike describes the attacker as “very likely a Chinese speaker after the profit motive” (moderate confidence).
💡 Why Should You Care?
A single actor wielding AI agents can hit many targets fast (per CrowdStrike). Defensive tools cut both ways — ARTEX’s developer has since gone fully closed-source — and regulators warn the leaked loan data could fuel phishing attacks.
🕵️ Hackers Are Selling Stolen AI Access at 97% Off
Brief Buzz:
Hackers want AI too — they just won’t pay for it. Google’s threat team reports a sharp rise this year in LLM-jacking: stealing AI accounts or compute to run cheap attacks against the very companies trying to stop them.
- Per Google researchers cited by the FT, dark-web sellers advertise access to stolen models from OpenAI, Anthropic and Google at discounts of up to 97%.
- Some sellers promise “guaranteed access” — free replacement logins if an AI lab blocks the stolen account.
- Prices for stolen AI accounts more than doubled in 2026 on rising demand for Claude and Gemini.
- Attackers also break into companies’ cloud servers and run their own models there, leaving victims to pay the compute bills.
- Google says a suspected China-linked espionage group ran its own model on hijacked cloud servers to dodge AI-provider monitoring.
💡 Why Should You Care?
Cheap, stolen AI hands attackers an edge over defenders paying full price. If your company is rolling out AI, watch for unexplained spikes in compute usage — and lock down your AI logins and API keys.
💸 Your AI Agent Needs a Spending Kill Switch
Brief Buzz:
AI agents can spin up paid services in minutes — and rack up bills while you sleep. Simon Willison (co-creator of Django) argues every pay-per-use service should ship with a default hard budget cap, on by default, that stops spending once it’s hit.
- Warning emails aren’t enough — a midnight alert can sit unread while a runaway service burns hundreds or thousands of dollars.
- Google Cloud’s Spend Caps hit public preview in July, capping one service per project (e.g. the Gemini API or Cloud Run).
- AWS followed in September with project spend limits that pause a project for the month — again, only for some new accounts.
- Willison wants removing a cap to be a deliberate checkbox, and agents to prefer providers that offer hard caps.
- Caps aren’t instant: the Firebase spend-cap docs note they aren’t true hard caps, so set yours well below your real limit.
💡 Why Should You Care?
If you let agents deploy code, an uncapped API key is a blank check. Before you ship, list every paid service without a hard cap and set one up wherever you can.
🔍 Stopped Reading AI Code? Build Checks You Can Trust
Brief Buzz:
AI now writes code faster than people can read it. In a recent essay on unread code, Google engineering leader Addy Osmani argues it’s fine to skip a line-by-line review in most cases — but something has to replace it. He admits he isn’t neutral on the question.
- Two things set the pace: checking code no one will read, and fixing the failures those checks miss.
- Anthropic says engineers marked under 1% of its AI reviewer’s findings as wrong, while PRs getting substantial comments rose from 16% to 54%.
- Osmani’s fix: have AI review every change first, with lighter human review for low-risk changes and owners scrutinizing core, sensitive code.
- A human still approves each merge, and the checks should come from an independent source — for example, a human-written spec.
- Skipping checks to keep up, Osmani warns, only defers the cost until it’s bigger and lands on customers.
💡 Why Should You Care?
Is it OK to ship AI-written code? The real question isn’t whether you read all of it, but which changes get human eyes. Anthropic’s Code Review documentation shows one approach (research preview for Team and Enterprise plans).
🧪 Your AI Sounded Smarter. Did It Actually Get Better?
Brief Buzz:
You tweaked the prompt and the replies sound better — but are they? In a new guide to testing AI systems, AI engineer Sergii Makarevych says the answer is evals: repeatable tests that score every change. His worked example is a support bot for a fictional bike shop.
- Of 60 test replies, 34 passed human review. Simple code checks caught missing facts but missed an invented refund timeline.
- Agreement between his AI grader and human graders was just 0.15 — well under the usual 0.6 threshold.
- Two prompts tied at 70%, yet the newer one was worse on 14% of cases. Averages can hide regressions.
- A 60-case set can’t reliably detect a change smaller than about 12 points — small test sets miss small drops.
- Leaderboards help shortlist models, but your own evals, with a clear release-blocking threshold, should decide what ships.
💡 Why Should You Care?
Build or buy, evals separate real progress from replies that merely sound better — and stop a bot from quietly degrading unnoticed. Hamel Husain’s evaluation guide, from a former GitHub engineer, shows the same idea on a real product.
🤖 Agility’s Factory-First Humanoids Head to Wall Street
Brief Buzz:
Agility Robotics built Digit, a two-legged robot already doing paid work in warehouses and factories — and it’s now heading for the public markets, betting on robots for industrial jobs rather than general-purpose home helpers.
- Agility is valued at $2.5 billion and will raise over $620 million via a merger with Michael Klein’s SPAC, listing under the ticker AGLT.
- Foxconn, which assembles the iPhone, led the $200 million private investment — a bet on mass-producing robots.
- Amazon, Schaeffler, GXO and Toyota already run nearly 100 Digit robots, logging more than 65,000 hours.
- CEO Peggy Johnson is focused on warehouse material handling for now; she says home robots come much later.
- Agility says the next-gen Digit v5 is its first robot designed to work safely alongside people without cages.
💡 Why Should You Care?
Humanoid robots are moving from demo reels to real paychecks. Agility’s public debut offers a rare early look at the actual economics of automating repetitive factory work.
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.
