One-Line Summary

Anthropic’s most powerful model leaks via a simple CMS mistake, while Google turns Search into a real-time voice-and-video assistant and OpenAI tightens its roadmap to what pays.

Big Tech

OpenAI Narrows to Revenue-Centered Features

OpenAI, maker of ChatGPT, pivots away from risky consumer plays—like an erotica feature, the Sora video app, and in-chat checkout—to double down on enterprise-friendly tools and monetization, a move that also aligns with a likely IPO timeline. The company still boasts about 900 million weekly active users and 50 million consumer subscribers, but is focusing on converting more of that demand into paid revenue. ¹

Behind the scenes, reporting points to a broader “pre-IPO cleanup,” including removal of 2–3% of training data tied to adult content and tightening on other sensitive domains (politics, medical, finance) to reduce regulatory and reputational risk. That kind of risk surface reduction is common before going public and helps enterprise sales cycles. ²

The strategic signal: OpenAI wants to be the default work copilot and agent platform, not an entertainment hub. With Apple reportedly opening Siri to rivals like Claude and Gemini, OpenAI’s path to consumers may look more like being the behind-the-scenes engine while it monetizes the business side where the willingness to pay is clearest. ¹

Google Search Live Goes Global

Google expands Search Live to 200+ countries and territories, letting users hold real-time, multimodal conversations with Search via voice and continuous live camera video, powered by the inherently multilingual Gemini 3.1 Flash Live model. This shifts search from “type and click” to “show and ask,” especially useful when you can’t name the thing in front of you. ³

Practically, you open the Google app, tap the Live icon, talk, and optionally point your camera at a leaky pipe, unknown equipment, or tangled cables; Search sees the context and responds in real time with guidance and links. It builds on Google Lens but keeps the video stream going like a FaceTime with the web, making troubleshooting and local discovery more natural. ⁴

The big commercial question: does this become a new distribution channel? If users engage in live, conversational flows with limited on-screen real estate, brands will need playbooks beyond classic SEO to surface products and services at the exact moment of intent. The rules for AI search placement are still being written. ³

Industry & Biz

Anthropic’s ‘Mythos/Capybara’ Model Leak Raises Cyber Stakes

A draft blog post left in an unsecured, publicly searchable data cache revealed Anthropic is testing “Mythos,” internally described as “by far the most powerful” model it has built, within a new, larger “Capybara” tier that reportedly beats Claude Opus 4.6 on coding, academic reasoning, and cybersecurity tasks. Anthropic confirmed early-access trials, called it a step change, and removed the cache after being contacted. ⁵

The draft framed “unprecedented” cybersecurity risks—directly relevant to blockchain and DeFi, where smarter AI can both discover vulnerabilities faster and automate exploitation. Recent episodes underscore the tension: Ripple’s AI-assisted red team found 10+ issues in XRP Ledger’s 13-year-old code; Ethereum launched a post-quantum security hub; and the Resolv stablecoin de-pegged after a contract flaw—exactly the kind of weak spot better AI could find or attack. ⁵

For decentralized AI, the leak resets the benchmark. Bittensor’s Covenant-72B pushed TAO up 90% and subnet tokens to a $1.47B combined cap, but a “centralized step change” widens the gap they must close. Anthropic says it is being deliberate about release given high run costs and safety risks—the irony being that a company hyping security left its own announcement in an unsecured cache due to human error. ⁵

Kandou AI Raises $225M to Extend Copper’s Life in AI Data Centers

Swiss interconnect company Kandou AI raises $225M at a $400M valuation, backed by SoftBank, Synopsys, Cadence, and Alchip, betting its Chord signaling can push copper links toward 448 Gbps+ with far lower power and cost—challenging the industry’s rapid pivot to photonics. Interconnect power can hit ~30% of cluster draw at 224 Gbps; cheaper, faster copper could delay optical overhauls. ⁶

Strategic investors matter: Synopsys and Cadence suggest potential IP/licensing routes into standard chip design flows, while Alchip offers a path to production. The Arm-like licensing model could scale faster than shipping standalone chips—if performance claims hold in hyperscale training clusters. ⁶

Skeptics note the ceiling: with models racing toward terabit-scale fabrics, optics promise a higher ultimate bandwidth/latency frontier. Kandou’s valuation—roughly one-tenth Ayar Labs—reflects that caution. The raise buys time to prove “good-enough copper” for another generation before a full optical turn. ⁷

Shield AI Raises $2B and Buys Aechelon to Accelerate AI Pilots

Defense AI company Shield AI secures $1.5B Series G at a $12.7B post-money valuation plus $500M in preferred equity (and a $250M delayed draw facility), and plans to acquire Aechelon Technology, known for high-fidelity simulation, physics-based sensors, and synthetic reality used in the Pentagon’s Joint Simulation Environment. Advent leads; JPMorganChase participates; Blackstone provides preferred equity. ⁸

The rationale: unify simulation and deployed autonomy to speed training and validation of Hivemind, Shield AI’s autonomy stack—already piloting 26 classes of vehicles from F-16s to UAVs. Aechelon will operate independently post-close, reporting to Shield AI’s CEO, to drive a foundation model for defense trained in sim and refined in the field. ⁹

Strategically, this is software-defined defense: develop and iterate AI pilots in high-fidelity digital twins, then harden with operational data. Capital lets Shield AI parallelize compute-hungry training and deployment, while funding next-gen platforms like X-BAT—positioning against a long shift to autonomous air power. ¹⁰

Community Pulse

Hacker News (80 upvotes) — Skepticism toward Google’s AI mediation of content, framed as paternalistic headline rewriting.

"I think this post sums it up pretty well: https://preslav.me/2026/03/25/murmel-competition-identity-tr..." — Hacker News

"And that sucks. I don't want the paternalistic thought policing." — Hacker News

What This Means for You

• Security leaders: The Anthropic leak is a wake-up call. If models can outpace human red teams on vulnerability discovery, your SDL needs AI in the loop—both for scanning your code/contracts and for simulating attacker behavior before mainnet or production pushes. Budget for model-enabled appsec, and rehearse incident playbooks where exploitation moves at machine speed. ⁵

• Marketers and product teams: Google’s Search Live could become a new “zero-click, zero-typing” channel. Think live troubleshooting, local services, and product discovery through video-and-voice. Start prototyping conversational answers that fit tiny visual surfaces and fast back-and-forth—an SEO-for-voice-and-video mindset. The early movers will define the playbook. ³

• CTOs and infra engineers: Kandou’s copper push is a pragmatic hedge if optics’ cost/complexity slows rollouts. If you’re capacity-constrained by interconnect power and reach, track licensing paths into mainstream EDA flows; a copper “bridge” could unlock one more GPU generation without a full photonics retrofit. But model growth argues for hybrid plans that keep terabit optics on the roadmap. ⁶

• Defense/robotics builders: Shield AI’s sim-first strategy is the template—train autonomy in rich digital twins, iterate faster, then validate in ops. Even in civilian robotics, that loop cuts time-to-reliability. Expect more capital to flow to simulation, sensor realism, and foundation models fine-tuned on operational feedback. ⁹

Action Items

1. Pilot Google Search Live use cases: On Android/iOS, test the Live icon with 3 customer scenarios (troubleshooting, local service discovery, product Q&A) and log what answers appear and how your brand shows up.
2. Add an AI red-team pass: Run an AI-assisted security sweep on one critical app or smart contract before your next release; compare findings and speed versus your manual process.
3. Run a conversational content sprint: Rewrite your top 10 support articles into tight, spoken, step-by-step answers that would work in a live, on-camera context.
4. Interconnect audit for 2026–27: Benchmark your cluster’s link power and reach; model costs for a copper-optimized upgrade versus optical pilots to de-risk next-gen training runs.