c/news

Microsoft's new quantum chip, Majorana 2, extends qubit lifetime to 20 seconds on average, reaching up to a minute in the best cases. This marks a 1000x stability improvement over last year's model.

The breakthrough comes from switching superconducting materials from aluminum to lead, which better shields qubits from external noise that causes errors.

Longer qubit lifetimes bring practical quantum computing closer, enabling complex tasks like drug discovery and logistics optimization. Microsoft now aims for a scalable quantum computer by 2029, halving its previous timeline.

📊@tech

According to Cloudflare Radar, 57.5% of all HTML page requests last week came from bots, including crawlers, AI data collectors, and scripts.

Only 42.5% were from real users in browsers.

JSON now accounts for 33.1% of all HTTP traffic, surpassing HTML at 12%.

JSON is used for machine-to-machine API calls, while HTML is what humans see in browsers.

📊@tech

ChatGPT apps can now perform real actions in connected apps, including sending and editing emails. Important actions require user approval and have built-in security checks.

This lets users send and edit emails directly within the chat interface. ChatGPT acts more like a full work assistant by handling confidential tasks safely.

📊@tech

By May, over 80% of the code in Anthropic's production base was written by Claude. In Q2, engineers merged eight times more code daily than in 2024, thanks to Claude automating much of the development work.

Claude is a model that improves itself and writes code to build its next version. Humans still lead in setting goals and handling complex context, but agents already outperform people in research tasks with clear metrics.

In one AI safety experiment, agents improved solutions by 97% in a week, while humans managed 23%. Anthropic sees this self-improving cycle accelerating faster than expected and calls for coordination tools, including possible pauses in frontier AI development.

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session. Dismiss alert {{ message }} anthropics / claude-code Public Notifications You must be signed in to change notification settings Fork 21.2k Star 131k [FEATURE] Official Claude Desktop build for Linux (Ubuntu LTS / Debian) #65697New issueCopy linkNew issueCopy linkOpenOpen[FEATURE] Official Claude Desktop build for Linux (Ubuntu LTS / Debian)#65697Copy linkLabelsarea:desktopenhancementNew feature or requestNew feature or requestplatform:linuxIssue specifically occurs on LinuxIssue specifically occurs on LinuxDescriptionpowell-clarkopened on Jun 5, 2026Issue body actionsPreflight Checklist I have searched existing requests and this feature hasn't been requested yet This is a single feature request (not multiple features) Problem Statement Preflight note. The closest open issue is #40347. Related: #47316 (closed), #38276 (closed as out of scope for this repo), #36011 (stale). I am filing this as a consolidation and extension of #40347 with corrected technical framing (Claude Code plugin development against Desktop extensions), named primary sourcing for the Cowork Linux-VM architecture, and current market data. Happy to merge into #40347 if maintainers prefer; please route rather than close if a different venue is correct.

On scope: this issue concerns Claude Code in two concrete ways. (1) Claude Code plugins are developed and tested against Claude Desktop extensions, which has no Linux build, so plugin work currently requires switching OS. (2) Cowork invokes the Claude Code binary inside a Linux VM on macOS, so the Linux execution path already exists inside the Claude Code product and is the practical thing missing as a published target.

What this issue is asking for A public Anthropic position on Linux desktop support, and ideally a first-party build. A reasoned "not on the current roadmap, and here is why" would resolve most of what this issue is about. There is, to my knowledge, no public statement on Linux desktop support; the absence is itself part of the problem.

Anthropic distributes Claude Desktop for macOS and Windows only. The official download page states "Not available for Linux". Claude Code (the CLI) runs natively on Linux but is a terminal tool, not a substitute for the desktop GUI. Desktop extensions (the surface Claude Code plugins are tested against), computer use, desktop dictation and Cowork are available only in Claude Desktop. Linux users therefore have no officially supported graphical path to these capabilities, and in particular no way to develop and test Claude Code plugins as desktop extensions without switching to macOS or Windows.

Why this is structurally hard to justify Anthropic already builds, signs and distributes Linux software. Per code.claude.com/docs/en/setup, Claude Code ships signed apt, dnf and apk repositories and per-architecture binaries (linux-x64, linux-arm64, musl variants). The pipeline exists.

The Cowork agent already depends on Linux inside the product. Independent reverse-engineering by Simon Willison on launch day (12/01/2026), corroborated by Pluto Security and pvieito ("Inside Claude Cowork"), found that on macOS Cowork boots a custom Ubuntu 22.04 VM via Apple's Virtualization Framework (VZVirtualMachine) and runs the Claude Code binary inside it under bubblewrap and seccomp. Anthropic's own documentation confirms the hypervisor split: Apple Virtualization.framework on macOS, Hyper-V on Windows. The community project johnzfitch/claude-cowork-linux demonstrates the same Cowork mode running natively on Linux x86_64 by stubbing the macOS native modules and skipping the VM entirely. The Linux capability already exists inside the product; what's missing is a published Linux target.

Source: github.com

I like Hacker News, but I do not love that so much of it has turned into discussion of a single topic: AI. So, I've created my own version.

HNSansAI pulls from the official Hacker News API, looks through the content of each article, and discards any that mention a term related to "AI". That includes stuff like:

I just put my head in the sand, works every time/

Does this use an LLM to categorize "AI-related" vs "not-AI related" articles? Would be ironic. Lol

RSS is right there @div https://elijahpotter.dev/rss.xml

noice. i am as tired of AI as the AI whiners reminds of the early cloud era and all the grey beard server admins with closests of servers in some strip mall office.... "NO COMPANY WILL STORE THEIR DATA IN THE CLOUD THEY DON"T CONTROOOOL" how did that turn out?

Source: elijahpotter.dev

Meta is notifying thousands of people whose Instagram accounts were hijacked during the months-long abuse of the company's AI chatbot, which hackers repeatedly tricked into taking control of a person's account.

In a new data breach notification letter, seen by this week in security, Meta has revealed for the first time how many people had their accounts hijacked as part of the long-running hacking campaign, which was discovered earlier this week and first reported by 404 Media ($) and TechCrunch ($). The number of affected accounts gives some clarity as to how widespread this hacking campaign was, and for how long it operated.

According to the data breach notice filed with Maine's attorney general's office late on Friday, Meta notified at least 20,225 people that their accounts had been compromised, including 30 people in Maine.

The compromises allowed the hackers to take over the person's entire Instagram and any linked accounts, including obtaining contact information, dates of birth, and profile information, as well as the ability to access the person's posts, direct messages, and account activity, the notice reads.

Meta's notice confirmed that the breach relates to "a vulnerability in an AI-assisted account recovery system for Instagram," which was exploited to "perform password resets on Instagram user accounts."

As previously reported, hackers abused a flaw in Meta's chatbot that allowed anyone to reset the password of any account that did not have two-factor authentication switched on. The bug tricked the chatbot into sending a verification code to an email address controlled by the hacker, rather than the account holder's email address on file, simply by asking it. The chatbot complied anyway.

Source: this.weekinsecurity.com

In my predictions for 2030 I wrote that tech writers would be using specialized LLMs, running locally on powerful hardware. I see hints of this move to “local first” among engineering pundits, but we’re not there yet, in part because of how much more powerful connected frontier models are. That doesn’t mean we can’t experiment, though. That’s precisely what I did last week, trying to fine-tune an instruct model to write like a software technical writer from the 80s and 90s.

To train a personal, local model to write like a technical writer from the 90s, one needs tons of written sources. If I wanted to fine-tune a model to write like myself, for example, this blog would not be enough, as it’s barely 100k words at the time of this post. You would need more samples for thorough training, and those are not easy to come by, nor simple to produce. The only quick way is to use an existing corpus. Where could I get one?

Meet Bitsavers: it’s a website that collects and scans old computer manuals and brochures. It’s an incredibly valuable repository of computer history and ancient tech writing, with mirrors available everywhere. As I’m fond of Microsoft manuals from the 90s, I chose the Microsoft collection as the source of training materials. The collection contains out-of-print docs published between 1977 and 2005: more than 37 million words, covering old systems and SDKs.

I downloaded the OCR’d text files and cleaned the content from artifacts and clutter (like indices and frontmatter) using good old Python scripts. I then used a cheap and fast model through OpenRouter, gemma-4-26b, to classify each paragraph as either “keep” or “drop” based on its intelligibility. This second pass cost around 8 dollars. Even with this two-pass cleaning, though, training data retained noise that I discovered only later, but that was largely OK for my tests.

I split the sanitized text into training examples on paragraph and section boundaries, breaking at headings and keeping code blocks whole, with each chunk capped at around 512 tokens as per Claude advice. Each chunk was paired with a synthetic instruction drawn from templates. I ended up with 192,456 examples in JSONL format (one JSON object per line). I could have used a small model to also come up with better instructions and questions, but I’m an impatient person.

In an ideal world, I would have several millions of dollars lying around, ready to be burned creating my own LLM, Fabrice. Since I’m far from rich (I wouldn’t be writing this otherwise), the alternative to Fabrice is fine-tuning, which involves tweaking the “weights” of a model so that each token generated is conditioned by the training materials. I like to picture fine-tuning as slightly steering the trajectory of a massive iceberg using tugs; just a little, just to get the intended effect.

Source: passo.uno

Due to recent regulation changes (전기통신사업법), the South Korean government is requiring internet communities and forum owners to scan every user uploaded images and videos on their website, by AI. The hardware to run these AI models are also not provided by government, website owners have to buy datacenter grade Nvidia GPUs by themselves, putting financial pressure to small businesses and forums.

Websites will need to implement these hardware and software features, starting immediately from July 1st, which is just next month.

Here is the original image provided from Korean government, specifying the hardware requirements for AI models. I also added English translated image, made with nano banana (sorry for using ai for this…)

Post from Korean forum/news website(루리웹) owner, after listening briefing from Korean government, expressing how ridiculous the situation is:

This feels really dystopian, even for South Korean standards.

And this is why South Korea has been ranked only slightly better in terms of press freedom than their authoritarian counterpart North Korea.

Source: discuss.privacyguides.net

Open Code Review is an AI-powered code review CLI tool. It originated as Alibaba Group's internal official AI code review assistant — over the past two years, it has served tens of thousands of developers and identified millions of code defects. After thorough validation at massive scale, we incubated it into an open source project for the community. Simply configure a model endpoint to get started.

It reads Git diffs, sends changed files to a configurable LLM via an agent with tool-use capabilities, and generates structured review comments with line-level precision. The agent can read full file contents, search the codebase, inspect other changed files for context, and produce deep reviews — not just surface-level diff feedback.

The Problem with General-Purpose Agents If you've used general-purpose agents like Claude Code with Skills for code review, you've likely encountered these pain points:

The root cause: a purely language-driven architecture lacks hard constraints on the review process.

Open Code Review's core philosophy is to combine deterministic engineering with an agent, each handling what it does best.

Deterministic Engineering — Hard Constraints

Source: github.com

For most of AI’s history, humans drove every step in its development cycle. But at Anthropic, we are delegating a growing share of AI development to AI systems themselves, which is speeding up our work.

Taken far enough, and given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor. This is called recursive self-improvement. We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.

Using public benchmarks and previously unreported data from within Anthropic, The Anthropic Institute is showing that AI is already accelerating the development of AI systems. To take just one example: today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.

The technical trends discussed in this piece suggest that AI systems are going to become much more capable in coming years. These trends have huge implications. AI that can build itself would be a major development in the history of technology—one that could bring enormous good for the world in science, healthcare, and beyond. But full recursive self-improvement also might increase the risks of humans losing control over AI systems. If systems are capable of fully building their own successors, the ways we secure them, monitor them, and shape their behavior all grow much more important.

In the early days, work at Anthropic looked like work at any other tech company: people writing code and docs on laptops.

People used early chatbots to help with parts of the process, like generating short code snippets and copying the output into text editors.

Source: anthropic.com

A reference implementation for autonomous vulnerability discovery and remediation with Claude, based on our learnings from partnering with security teams at several organizations since launching Claude Mythos Preview. For a write up of these learnings along with best practices, see the accompanying blog post (also available in blog-post.md). For a lightweight SDK-only walkthrough of the same recon → find → triage → report → patch loop, see the companion cookbook.

This repo is not maintained and is not accepting contributions.

🔒 Want a managed option? Anthropic offers Claude Security, a hosted product that finds and fixes vulnerabilities in your source code across multiple projects. Claude Security scans your repository for vulnerabilities, applies a multi-stage verification pipeline to reduce false positives, and lets you manage findings through their lifecycle: triage, fix validation, and rapid fix generation.

This repository is an open-source reference implementation based on general best practices for finding vulnerabilities using Claude. You can use it to build your own vulnerability finding pipeline, customize the logic, and it can be used with whatever access you have to Claude APIs (including Bedrock, Vertex, or Azure).

Claude Code skills: /quickstart, /threat-model, /vuln-scan, /triage, /patch, /customize: interactive scoping, scanning, triage, and patching. Open this repo in Claude Code and run /quickstart to get oriented. harness/: the autonomous reference pipeline (recon → find → verify → report → patch), configured for finding C/C++ memory vulnerabilities using Docker and ASAN. This harness is a reference, not a product. The general shape, prompts, and sandboxing are reusable, but the harness will not work on every codebase out of the box. Run /customize to port it to your language, detector, or vuln class. ⚠️ Security: /quickstart, /threat-model, /vuln-scan, and /triage only read and write files. Running /patch on static findings (TRIAGE.json or VULN-FINDINGS.json) is likewise read- and write-only. /customize edits the harness code and runs validation commands. Any of these skills are safe to run unsandboxed, as long as you review and approve each tool use in Claude Code. The autonomous reference pipeline (including /patch on pipeline results) executes target code, so it refuses to run outside of a gVisor sandbox unless explicitly overridden. To get set up, run scripts/setup_sandbox.sh once, then invoke the pipeline via bin/vp-sandboxed. See docs/security.md and docs/agent-sandbox.md for more details.

git clone https://github.com/anthropics/defending-code-reference-harness cd defending-code-reference-harness claude # 30-sec intro + guided first run on the canary target > /quickstart > /quickstart how do I port the pipeline to Java? > /quickstart how do I triage all these bugs? Further Reading Blog Post · The accompanying blog post with learnings + best practices Pipeline · How it works: diagram, stages, CLI flags Security · Sandboxing, what not to mount Agent sandbox · gVisor isolation + egress allowlist for every agent Customize · Port to my stack; which files change and why Patching · Generate and verify fixes for verified crashes Troubleshooting · Duplicates, rate limits, subagent model pinning Safeguards · Block for dangerous cyber work Ramp Up The most successful security teams we've partnered with are those that have gotten hands-on the fastest. Though it's tempting to spend months designing the perfect pipeline, we recommend starting small on Day 1 and building from there as learnings come. The steps below follow that pattern and set an ambitious (but reasonable) pace based on what we've seen.

Source: github.com

S&P Dow Jones Indices has decided to keep its existing rules for adding newly public companies to the S&P 500 unchanged, rejecting a proposed fast-track mechanism that would have allowed mega-cap IPOs like SpaceX to enter the index almost immediately after listing.

Under the current rules, companies must wait at least 12 months after their IPO and demonstrate sustained profitability before becoming eligible for inclusion. They also need to meet a 50% public float requirement — a threshold that could keep SpaceX out for years given its plans to list only about 4% of shares publicly.

The decision comes after a consultation period that attracted significant attention from investors and media. S&P ultimately sided with index stability over pressure to accommodate the wave of anticipated mega-IPOs from companies like SpaceX, Stripe, and OpenAI.

The ruling means index funds and ETFs that track the S&P 500 will not be forced to buy into these companies at IPO premiums, protecting passive investors from early-stage volatility. Individual investors can still trade these stocks through retail platforms once they go public.

Source: bloomberg.com

RAM price tracking through 2026 will show you that kits that routinely cost less than $100 just a year ago are now fetching upwards of $240 (16GB). As the AI frenzy has taken hold, retailers far and wide have been pumping up their RAM prices to exorbitant levels. However, there's so much fluctuation and noise that average pricing is now something of a ludicrous fugazi. The going rate for 32GB of DDR5 RAM - the cheapest you can expect to pay - has hovered around $320 for some time, climbing past $350 in recent weeks. Price tracking courtesy of PCPartPicker now reveals the cheapest 32GB DDR5 RAM you can buy is $375. Specifically, four XPOWER kits from Silicon Power will set you back $374.97 thanks to a promo code. You can see the listings yourself below.

As you can imagine, this is enormous pricing pressure for enthusiasts trying to build gaming PCs or upgrade their rigs in 2026. A component that once cost less than $100 and was something of an afterthought now costs almost four times as much, and that's before you've even fired a neuron in consideration of aesthetics, timings, or brand. More popular kits from the likes of Corsair and Crucial, or RGB offerings to match the rest of your build, will easily set you back more than $400.

Of course, 32GB is really the minimum sweet spot you should be aiming for when building a PC in 2026. If you did want more capacity, 64GB will set you back an astonishing $679.99. 16GB of RAM as a compromise can be found for $200 at B&H Photo, but with SK hynix warning that manufacturing constraints will persist through 2030, there's no sign of prices letting up so that you can upgrade capacity any time soon.

The humble RAM combo deals we've been highlighting in recent months are a small source of solace for builders, letting you score RAM for less than the $375 going rate if you pair it with a decent motherboard, a processor, or even an entire set of PC components. A theme of ongoing Computex 2026 announcements remains a lack of pricing clarity on lots of PC hardware, including Nvidia's RTX Spark laptops and PCs, as well as new-build systems and, of course, RAM components themselves. Vendors are likely wary of scaring off potential buyers with higher-than-expected prices ahead of release. Perhaps more likely, the prices haven't been set because they're still going up. Storage isn't much better, with SSD price tracking revealing that drives which once cost as little as $38 are now fetching $200.

AMD is making a noticeable effort to keep PC gaming prices down, this week announcing the return of its Ryzen 7 5800X3D, and the advent of a new Ryzen 7 7700X3D. Intel, which warned this week that "something has to give" when it comes to memory prices, also teased dragging out some of its legacy products to give users more options on older memory technologies, namely Raptor Lake and DDR4.

Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.

Source: tomshardware.com

The percentage of failing grades in multiple UC Berkeley computer science classes in spring 2026 is significantly higher than past semesters and marks a departure from the department’s grading guidelines.

According to Berkeleytime, 35.3% of CS 10 students and 10.6% of CS 61A students received F’s in spring 2026.

The percentage of failing grades in multiple UC Berkeley computer science classes in spring 2026 is significantly higher than past semesters and marks a departure from the department’s grading guidelines.

Instructors point to students’ increased reliance on AI, lack of mathematical preparedness and understaffing as potential contributing factors.

According to Berkeleytime, 35.3% of CS 10 students and 10.6% of CS 61A students received F’s in spring 2026. In spring 2025 and spring 2024, the percentage of F’s did not exceed 10% for either class. The electrical engineering and computer sciences department’s grading guidelines state that 7% of students in lower division courses, including CS 10 and CS 61A, should receive D’s and F’s.

In addition, the guidelines state that “a typical GPA for a lower division course will fall in the range 2.8 - 3.3.” In spring 2026, both classes’ average grades were C-pluses, according to Berkeleytime, corresponding to a 2.3 GPA.

Source: dailycal.org

Uber Caps Usage of AI Tools Like Claude Code to Manage Costs. I wrote the other day about Uber blowing its 2026 AI budget in four months, and how that wasn't particularly surprising given they would have set that budget in 2025, before anyone could have predicted how popular token-burning coding agents were about to become.

The rideshare giant is limiting all employees to $1,500 in monthly token spending per AI coding tool, an Uber spokesperson said in response to a Bloomberg News inquiry. That means spending on one tool doesn’t have a bearing on the budget for another. The limits, which have been instituted in recent months, only apply to agentic coding software such as Cursor or Anthropic PBC’s Claude Code.

A $1,500 monthly limit per tool strikes me as a rational policy response to over-spending, and much more sensible than those tokenmaxxing leaderboards encouraging employees to compete for as much AI usage as possible.

It's also interesting in that it hints at a real dollar value for what Uber is getting out of these tools. If we assume two actively used tools per engineer that's $3,000 * 12 = $36,000 cap per engineer per year. Levels.fyi lists the median yearly compensation package for Uber software engineers in the USA at $330,000.

That means each employee's AI spending cap is ~11% of that median compensation package.

I noted that my own token usage comes to about $1,000/month against each of Anthropic and OpenAI - which currently costs me just $100 per provider thanks to their generous subsidized plans for individual subscribers. Those plans are no longer available to larger companies like Uber.

Source: simonwillison.net

Edinburgh-based Wordsmith closed a $70 million Series B led by Highland Europe and Index Ventures, bringing total funding to $100 million. The platform currently serves over 500 in-house legal teams including BT, Canva, Starling, and Sage.

Most legal AI startups go after law firms where billable hours drive the revenue model. Wordsmith takes a different approach - it targets in-house legal departments, helping companies handle legal requests faster without leaning on outside counsel.

The platform works as a single entry point for legal requests across an organization. Requests come in through email, Slack, Salesforce, or Teams. The system reads them, assigns ownership and priority, applies existing playbooks to handle routine work automatically, and escalates anything that needs actual legal judgment. Everything gets logged.

CEO Ross McNairn put it bluntly: "Legal does not need another filing cabinet, and it does not need another copilot that simply helps one lawyer work faster. Wordsmith is the front door that does the work."

McNairn previously sold Dorsai Travel to Skyscanner and scaled TravelPerk to $170M ARR as CPTO. CTO Volodymyr Giginiak spent over a decade at Facebook and Instagram. The team is about 110 people across Edinburgh, New York, and London with plans to hit 300 by end of 2026.

The competitive field is getting crowded. Harvey raised $200M at an $11B valuation in March targeting law firms and enterprise legal research. Swedish Legora raised $600M at $5.6B going hard across European markets. Wordsmith is betting that organizational legal workflow - routing, triaging, resolving requests at the company level - is a different category than helping individual lawyers research faster.

Quantinuum is increasing the size of its initial public offering as investor demand keeps building. The company now plans to offer 26.5 million shares at $53 to $55 each, up from the original plan of 21 million shares at $45 to $50. At the top of the range it could raise $1.46 billion and hit a $14.3 billion market cap.

The revised terms are a significant jump from the initial filing in May, when Quantinuum was targeting roughly $1.05 billion at a $12.7 billion valuation. Shares are expected to start trading on the Nasdaq under "QNT" this week.

This would be the largest public listing ever by a dedicated full-stack quantum computing company, and the first pure-play quantum hardware firm to list on a major US exchange through a traditional IPO rather than a SPAC. For context, IonQ currently trades at around $27 billion and D-Wave sits near $11 billion.

Quantinuum was created in 2021 through a merger of Honeywell Quantum Solutions and Cambridge Quantum. It reported $30.9 million in revenue during 2025, up from $23 million the year before, along with $79.3 million in bookings. Net loss was $192.6 million as the company keeps investing heavily in R&D.

The company also has a tentative $100 million agreement with the US Commerce Department to advance its trapped-ion quantum systems. IBM is expected to receive about $1 billion from the same department for a standalone quantum foundry. After the offering, Honeywell will retain roughly 49% of voting power and Cambridge Quantum about 32.5%. Founder Ilyas Khan's personal stake would be worth over $2 billion at the offering price.

Findigs, an AI-powered rental screening platform, secured $32 million in Series C funding. The round was led by Marc Weiser and Adam Boyden at RPM Ventures, with participation from existing investors Nyca Partners, Frontier Venture Capital and Western Technology Investment. Total funding now stands at $80 million.

The company was founded in 2018 as a tenant screening business. Co-founder Steve Carroll says it's evolved: "Over the last two years, what we've built is an autonomous decisioning platform." Rather than just providing data for human review, the system now automates the approval decision itself.

Findigs calls what they do "leasing decisioning" - optimizing revenue through three levers: maintaining full occupancy, ensuring reliable rent collection, and cutting acquisition costs. The platform serves 400,000 units across hundreds of operators. Customers report up to 80% fewer evictions and 90% lower delinquency rates. McKinley, one of their bigger clients, hit 46% lower eviction rates, 33% reduced acquisition costs, and 98.6% occupancy last year.

New capital goes toward expanding affordable housing capabilities - LIHTC and Section 8 workflow support - plus launching rent guarantee products that cover operator revenue across the full lease term.

Hugh Frater, a BlackRock founding partner and former Fannie Mae CEO, is joining the board. His take: "The best way to manage risk is usually by not taking it. Using AI, Findigs combines better risk decisions with improved customer experience."

A decade ago defense tech was considered a niche, even controversial corner of venture capital. Few startup investors dared to bet on companies working with the military.

That's changed fast. More than $14.6 billion in venture investment has gone into defense, national security and law enforcement startups through the first five months of 2026, blowing past the sector's previous annual record of $9.6 billion raised in all of 2025.

The rise has been building for years. Global defense tech funding totaled $1.6 billion in 2020 before climbing to $3.9 billion in 2021. Funding then held relatively steady between $2.8 billion and $3.8 billion from 2022 through 2024. That changed last year when it jumped to $9.6 billion. Now five months into 2026, the full-year record is already broken.

The biggest contributor by far is Anduril Industries, which closed a $5 billion Series H at a $61 billion valuation. But it's not alone - Shield AI raised $2 billion, Saronic secured $1.75 billion for autonomous naval vessels, and Mach Industries announced a $300 million Series C today at a $1.8 billion valuation. Deal flow has stayed steadier while check sizes balloon - 107 rounds so far versus 206 in all of 2025. The capital is concentrating into fewer, bigger bets.

Investors are starting to eye exits too. AI drone company Swarmer went public this year and its stock soared over 500% on its first day of trading. Anduril is widely viewed as the most likely IPO candidate - a public offering at that scale would be the first real test of public-market appetite for next-gen defense contractors.

Anthropic submitted a confidential IPO registration with the SEC, putting the company on track for a public listing at a valuation close to $1 trillion. The filing comes as the company scales Claude and pushes deeper into enterprise contracts.

The move sets up a direct race with OpenAI, which is chasing its own public offering. Both companies need billions to fund compute infrastructure and talent. Pricing and timeline haven't been disclosed yet.

The bigger question is whether the market can absorb what's coming. SpaceX, OpenAI, and Anthropic are all targeting IPOs in a similar window. Together the three could put roughly $200 billion in new shares on the market. For context, total US IPO activity was $44 billion in all of 2025.

If the listings succeed they'll open the floodgates for other AI startups looking to cash out. If they don't, the damage won't stay contained to AI - over a third of the S&P 500 already has significant exposure to AI-heavy tech stocks through indirect holdings.

Alphabet announced it's raising $80 billion through equity offerings to fund AI infrastructure expansion. The company says demand for its AI services is exceeding available supply.

The deal breaks into three pieces: $30 billion in underwritten public offerings, $40 billion in an at-the-market program starting Q3 2026, and a $10 billion private placement to Berkshire Hathaway.

Google's 2026 capex guidance stands at $180 billion to $190 billion, roughly double its 2025 spend of $91.4 billion. CEO Sundar Pichai shared the numbers at Google I/O last month. Across all Big Four hyperscalers, combined AI capex is expected to pass $700 billion this year, with estimates pointing toward $1 trillion by 2027.

Sekai raised $20 million in a Series A led by Khosla Ventures and Connect Ventures. Additional investors include a16z Speedrun, Mayfield, A*, MVP Ventures, 359 Capital, Parable VC and 645 Ventures. Total funding is now around $30 million.

The app lets people create mini applications by describing what they want in a text prompt. Users can also remix and modify apps that others built. Examples range from outfit-selection tools based on weather and mood to personality quizzes and interactive birthday cards.

The numbers stand out: over 15 million mini apps created, more than 200,000 generated daily, and average daily usage exceeding one hour per user. Early traction came from shareable creations - particularly interactive birthday cards that spread virally.

Founder Lucky Zhang positions Sekai as the opposite of doomscrolling: "The way to get rid of doomscrolling is asking you to interact." This is his fourth startup - Apple acquired his video e-commerce company Yi+ AI in 2017, ByteDance bought his Latin American short-video platform Blacktail in 2020.

Sekai originally focused on anime fan-fiction experiences before pivoting to broader app creation as AI coding models improved. Connect Ventures' Nicole Quinn explained her investment: "The fact that customers are spending over an hour a day on Sekai is the kind of signal that there's true obsession with a product."

The company is now hiring engineering and product, and exploring creator monetization through apps and interactive fan experiences.

Israeli networking company DriveNets raised $410 million in a Series D round at a valuation of $8.5 billion. Total investment in the company since its founding now stands at roughly $1 billion.

DriveNets builds networking solutions for AI infrastructure - specifically the Ethernet fabric layer that connects thousands of processors inside AI data centers. The technology uses standard hardware combined with proprietary software that enables high-performance scaling across massive compute clusters.

The company says it's cash-flow positive with more than $1 billion in orders and project backlog. It's also exploring a potential secondary transaction later this year.

The round was led by Bessemer Venture Partners and Atreides Management, with AMD joining as a strategic investor alongside existing backers Pitango and D1 Capital Partners.

DriveNets was founded by Ido Susan and Hillel Kobrinsky. Susan previously founded Intucell, which Cisco acquired in 2013 for $475 million after raising just $6 million. Last July AT&T bought shares from DriveNets employees and investors for $650 million - the telecom giant has been a major customer, deploying DriveNets' Network Cloud across its core network.

The less visible part of the AI infrastructure stack - networking between GPUs, not the GPUs themselves - is turning into a massive market. As clusters scale to tens of thousands of accelerators, how fast data moves between them becomes the bottleneck, not compute power alone.

Business insurance startup Corgi closed a $160 million Series B led by TCV, valuing it at $1.3 billion. That's just four months after the company announced a $108 million Series A. Total raised is now $268 million.

Nico Laqua and Emily Yuan started Corgi in 2024 and went through YC's Spring 2024 batch. The company is a full-stack insurance carrier covering general liability, cyber liability, and tech/AI liability. Customers include Deel and Artisan. Other investors in the round: Kindred Ventures, Leblon Capital, and First Order Fund.

Corgi handles underwriting, policy management, and claims through its own infrastructure instead of relying on third-party insurers. It's now expanding from startup insurance into trucking.