Claude Lies to Police, AIs Nobody Wants, & AI-led Hacks

TLDR;

Welcome back! This week we have a lot of news and insights. A single guy hacking several South Korean banks using AIs, another “naughty naughty” moment from an AI model actually lying to the police, an AI product that, in my personal view, should never exist, the latest market and M&A data, and new products including the latest sandboxed agent, the new hottest trend in AI, coming from Google, and many other insights to keep you abreast of the industry.

Enjoy!

CYBERSECURITY

Random Guy Hacks South Korean Banks Using AI Agents

An unidentified attacker used AI tools to steal data from South Korean financial organizations between late September and early October, according to a CrowdStrike investigation published October 7.

The campaign combined ARTEX, a Chinese-developed tool that uses AI agents to test systems for vulnerabilities, with large language models. Interestingly, while the user leveraged a wide range of models, including both Chinese and American (e.g., Grok), they used Claude Code as the harness. Researchers recovered exposed Claude Code session histories, configuration files, and AI memory files from attacker-controlled servers, so they likely got caught.

ARTEX primarily used DeepSeek v4.1-flash, while additional Claude Code sessions used GLM-5.3 and Grok 4.6. The records also showed requests for help finding places to sell stolen Korean data, including Telegram groups.

The company assessed with moderate confidence that the operator was likely Chinese-speaking and financially motivated; it did not attribute the campaign to a named group.

TheWhiteBox’s takeaway:

Hilariously, one session contained a request to write a security researcher résumé presenting results from the same campaign as professional achievements. The kid, who’s only 22, has guts, but left too many footprints at the scene of the crime.

Which raises the question: can anyone hack anything with AI?

The answer is obviously no; this guy/gals was inexperienced but certainly knowledgeable.

However, it’s undeniable that these tools can empower both good and bad people, which means that it seems we’re entering a new world of cybersecurity attacks on another scale.

Needless to say, the same company reporting this, CrowdStrike, and others like Palo Alto Networks are creating defense swarms to fight against AI attacks, which may not be more sophisticated than human ones, but are executed at a scale that makes them really hard to fight against without AIs of your own. In a way, it’s like going into a gunfight with a knife.

Of course, the immediate urge is to ban open models, which are more likely to be used (in theory, at least) for cyberattacks.

But that view has two big issues:

If only private models are allowed, whether you can defend yourself from Claude depends on you having access to your “defense Claude.” In that world, Anthropic decides who gets hacked and who doesn’t. If we’re trying to create an Anthropic/OpenAI world duopoly, this is sure as hell one great way to promote it.

The best defense against these events is precisely the opposite of an open-source ban. In fact, you want open-source…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.