Covering stories collected through the morning of November 7 (Japan time), it was a day when examples of AI being used as an "attack-side" tool turned abruptly concrete, even as opposition to AI infrastructure surfaced in the form of physical action: the leak of all customer data from security firm Chainalysis, the hacking capabilities of free Chinese models, and a protest break-in at an Amsterdam data center. On the market side, too, the signals of slowing AI spending kept stacking up.

Chainalysis Loses All Customer Data in an "AI-Agent-Assisted" Attack

Chainalysis, the blockchain analytics giant, revealed on November 6 that a breach of its network allowed hackers to obtain sensitive information on every user: names, usernames, password hashes, and currently valid tokens. According to Ars Technica, the company described it as a "cyberattack from multiple fronts," saying the attack combined techniques that used AI agents for assistance with methods that siphon data off in transit.

The incident was first disclosed in late October. At the time, the company said that "non-sensitive data such as user IDs, names, email addresses, and roles within their organizations" had been accessed — but the scope has now effectively grown to "the credentials of all customers." Outside attackers are said to have used AI-agent-assisted attack techniques to gain unauthorized access to the company's systems and internal tools.

What makes this incident especially serious is Chainalysis's customer base. Public records confirm that its clients include government agencies such as the FBI, U.S. Immigration and Customs Enforcement (ICE), and the U.S. Air Force, as well as private companies like Microsoft, Coinbase, and Barclays. Whether the leaked data will lead to secondary harm remains unknown at this point, but the fact that the very toolbox of blockchain "surveillance" itself became the target is a symbolic turn of events.

Free Chinese AI Models' Hacking Capabilities Close In on OpenAI's Most Advanced Models

In an experiment that tasked AI models with breaking into simulated networks, freely available Chinese models proved not that far behind OpenAI's most advanced, cybersecurity-specialized (and unreleased) model. In the experiment, reported by WIRED, models from Chinese players such as DeepSeek and GLM posted scores approaching those of OpenAI's unreleased, state-of-the-art cyber model.

OpenAI offers its cyber-offense-focused models only to a restricted set of customers, but if Chinese models that anyone can pick up for free are closing in on that level, the threat landscape widens at a stroke — the barrier to entry for attackers could drop to little more than "an API key and prompts in English." And as the reported use of AI agents in the Chainalysis attack above suggests, AI-powered attacks are no longer a story about the future; they have entered the stage of being cited as an explanation of the present.

Protesters Break Into Microsoft Data Center Site in Amsterdam

Protesters forced their way into a Microsoft AI data center facility in Amsterdam, the Netherlands. Al Jazeera and Bloomberg reported the incident, and video footage of the break-in is circulating. No injuries or major damage to equipment have been confirmed at this point.

Small as it may be as a single incident, the context is worth noting. In this week's U.S. midterm elections, 40 of 41 data-center-related ballot measures came down on the side of tighter restrictions on development (as covered in this week's roundup). The local backlash against the enormous electricity and water consumption of AI's computing base can be read as changing shape, from votes to physical protest. And as Google's announcement of a data center investment in Wyoming — just two days after those restriction votes — showed, the gap between corporate investment plans and community consensus looks set to keep seeding friction in region after region.

Perplexity Launches "Teammates," an Always-On Agent — a "Colleague" You Hire for $200 a Month

Perplexity, the search-focused AI company, has launched "Teammates," agents that stay up and running at all times. According to The Decoder, each Teammate runs on its own virtual machine and is designed to keep working through ongoing projects rather than responding in discrete turns.

Access requires the $200-per-month Max plan, a fairly steep price even compared with similar services from OpenAI and Anthropic. But the very category of "an agent that stays resident and keeps pushing work forward" is only now taking shape, so the pricing will likely be judged once the offerings flesh out. Whether the three-layer split — conversations like ChatGPT, agents that act only when instructed, and agents that run continuously — settles into place will be a major theme going forward for both UIs and billing models.

Microsoft to Restructure Its neocloud Relationships — Anthropic's IPO Also Draws "$1.1 Trillion" Talk

On both the infrastructure and capital-markets fronts, reports pointing to a slowdown in the AI buildout piled up.

According to CNBC, Microsoft plans to restructure its contractual relationships with neoclouds — the emerging cloud providers from which it procures AI computing capacity — and narrow down its list of partners. The report sees signs of "strain" in AI capital spending, suggesting that the capacity race, until now a story of one-way expansion, may be entering a phase of rationalization.

On the capital-markets side, Bloomberg reported that ahead of Anthropic's IPO marketing officially kicking off next week, buyers are discussing a valuation range of $1.2 trillion to $1.3 trillion, with some arguing it should be brought down to as low as $1.1 trillion. Earlier reporting had investors resisting a $1.5 trillion valuation, and after the post-election plunge in AI stocks, the assumed range is taking concrete shape.

Meanwhile, an analysis by Ramp, which works with financial data, shows that while AI usage keeps climbing, spending has plateaued — the result of model price cuts and cost management by large customers. TechCrunch points out that this kind of trend could rattle the investor-side assumption that "AI hypergrowth will continue."

EU's New Rules May Be More Than AI Text Watermarks Can Deliver

IEEE Spectrum published an argument that current watermarking techniques may be unable to satisfy the new rules the EU is introducing for AI-generated text. In an evaluation that took up major watermarking techniques, including OpenAI's TextGrain, they can fall short of the level the rules demand in terms of detection reliability and resistance to tampering. Whether the provenance of generated text can keep resting on "watermarks" — the piece throws into sharp relief a mismatch in understanding between regulators and the technical side.

Other Developments

Security researchers discovered that hundreds of servers hosting the penetration testing tool "Artex" are exposed to the internet. That means anyone could get at the hacking tools, putting the issue of poorly managed tools leaking out back on the table (TechCrunch).

Underdog has released, in a new demo, an audio model of an extremely small 32MB. Despite its size, it can reportedly generate both speech and music, likely lowering the bar for audio AI on edge devices even further (TechCrunch).

AI turned to attack, and society pushing back against AI. Two currents that look like opposites — yet both are consequences of AI being built into real-world social systems. The next stage of the debate may be less about performance or features, and more about the design of trust and consensus.