AI’s New Security Problem

Today we will see, Anthropic just accused three major AI labs of industrial-scale model extraction. Over 16 million exchanges. 24,000 fraudulent accounts. Not hacking the weights, copying the behavior. This is the invisible war inside AI. Also, Guide Labs just open-sourced an 8B model that can explain every token it generates. Not after training. During training. Instead of probing a black box, they redesigned the box. That changes the future of AI trust.

In today’s post:

• The invisible war inside AI

• The model that explains itself

• $12 billion says AI isn’t slowing down

SPONSORED BY

Close more deals, fast.

When your deal pipeline actually works, nothing slips through the cracks. HubSpot Smart CRM uses AI to track every stage automatically, so you always know where to focus.

Simplify your pipeline with:

• Instant visibility into bottlenecks before they cost you revenue

• Clear dashboards highlighting deals in need of the most attention

• Automatic tracking so your team never misses a follow-up

Start free today. No credit card required.

Get started free

What’s Trending Today

REPORT

Distillation isn’t innovation. It’s extraction

Anthropic just published a report on large-scale distillation attacks targeting Claude. Three major labs allegedly ran over 16 million exchanges to extract model capabilities. Five sentences is all it takes to understand what’s happening. Distillation is normal. Every lab uses it. But there’s a line between internal optimization and external extraction. Cross that line, and it stops being research. It becomes industrial copying.

Here’s everything you need to know:

• DeepSeek, Moonshot, and MiniMax allegedly used 24,000 fraudulent accounts to generate millions of coordinated prompts designed to extract reasoning, coding, and tool-use capabilities.

• The traffic patterns weren’t organic, because they were synchronized, repetitive, and concentrated on Claude’s most differentiated strengths.

• One technique reportedly asked Claude to reconstruct its internal reasoning step by step, effectively mass-producing training data.

• Proxy “hydra” networks allegedly rotated thousands of accounts so that when one was banned, another immediately replaced it.

• MiniMax pivoted within 24 hours of a new Claude release, redirecting traffic to capture updated capabilities in real time.

• Anthropic argues that distilled models may strip out safeguards, which raises national security concerns beyond simple competitive advantage.

• The company is responding with behavioral fingerprinting, coordinated intelligence sharing, stricter access controls, and model-level countermeasures.

This is where it gets uncomfortable. If progress looks fast, we assume innovation. But what if some of that speed is borrowed? Distillation attacks complicate export controls because they allow capability transfer without transferring the original infrastructure. AI competition is shifting from model building to capability protection.
The next frontier isn’t just better models. It’s defensibility. If frontier labs can’t protect outputs, scale alone won’t matter. And if safeguards don’t transfer with capability, the risk surface widens.

RESEARCH

What if every AI answer came with receipts?

Guide Labs just open-sourced Steerling-8B. It’s an 8B model built to be interpretable by design. Most AI models feel like black boxes. You give them a prompt. They give you confidence. But you never see the wiring. And that uncertainty quietly erodes trust. Guide Labs is trying to flip that.

Here’s everything you need to know:

• CEO Julius Adebayo says every token the model produces can be traced back to its training origins.

• Instead of reverse-engineering behavior after training, they insert a “concept layer” during training that buckets data into traceable categories.

• That upfront annotation makes the model more structured, but also more controllable.

• The approach allows developers to toggle concepts on or off, like controlling how gender or violence is represented.

• The model reportedly achieves about 90% of frontier performance, while using less training data.

• The team tracks “discovered concepts,” meaning the model still forms new abstractions like quantum computing.

• Guide Labs argues this architecture is critical for finance, science, and any regulated industry where reasoning must be auditable.

This changes the frame. Most interpretability research studies models like neuroscientists. Probe. Measure. Guess. Hope the explanation sticks. Guide Labs is treating interpretability like engineering. Design first. Explain later. That’s a philosophical shift. The real question isn’t whether this model matches the best ones.
It’s whether the future will demand that all of them explain themselves

UPGRADE

Amazon just doubled down on the infrastructure race

Amazon announced a $12 billion investment in Louisiana data centers. This is part of a projected $200 billion capex year. For months, investors have been nervous. Margins matter. Efficiency matters. AI spending looks endless. But Amazon isn’t tapping the brakes. It’s pressing harder.

Here’s everything you need to know:

• Amazon will build new AI data center campuses in Caddo and Bossier Parishes, expanding its cloud and AI footprint in the U.S.

• The company expects 540 full-time jobs and roughly 1,700 additional roles tied to construction and operations.

• Amazon’s total capital expenditures could hit $200 billion this year, more than any other hyperscaler.

• Wall Street reacted skeptically after earnings, wiping hundreds of billions off Amazon’s market value.

• Most of the spending will fund AI infrastructure, including chips, networking, and large-scale data centers.

• The company says it will cover 100% of local infrastructure costs and invest up to $400 million in water systems.

• Cooling will rely partly on natural air, and Amazon claims only surplus water will be used.

Tech giants are racing to secure power, land, and grid capacity before someone else does. AI models don’t just need better algorithms. They need electricity. And electricity is becoming strategic. The AI race has quietly become a construction race. Whoever controls compute controls capability. Whoever locks in energy controls compute. Short-term investors see rising capex and shrinking free cash flow. Long-term strategists see defensive positioning.

Free Guides

My Free Guides to Download:

🚀 Founders & AI Builders, Listen up!

If you’ve built an AI tool, here’s an opportunity to gain serious visibility.

Nextool AI is a leading tools aggregator that offers:

• 500k+ page views and a rapidly growing audience.

• Exposure to developers, entrepreneurs, and tech enthusiasts actively searching for innovative tools.

• A spot in a curated list of cutting-edge AI tools, trusted by the community.

• Increased traffic, users, and brand recognition for your tool.

Take the next step to grow your tool’s reach and impact.

Reach 150,000+ READERS:

Expand your reach and boost your brand’s visibility!

Partner with Nextool AI to showcase your product or service to 140,000+ engaged subscribers, including entrepreneurs, tech enthusiasts, developers, and industry leaders.

Ready to make an impact? Visit our sponsorship website to explore sponsorship opportunities and learn more!