Probably the most vital rigidity on this challenge is between two firms making totally different selections about find out how to deal with AI with frontier safety capabilities. Anthropic restricted Claude Mythos to a small company cohort by Challenge Glasswing. OpenAI launched GPT-5.5 to normal availability, and a few are calling it “Mythos-like hacking, open to all.” The AI Safety Institute’s analysis confirms the potential is actual and consequential. How will you handle threat when the time between discovery of a vulnerability and exploitation collapses to zero?
One other essential theme is that, within the phrases of The Sequence, “AI is changing into operational.” It’s not about LLMs that may play video games with phrases. It’s about instruments that may automate processes throughout an enterprise: brokers, in fact, however extra particularly brokers that may be shared by groups to provide a constant set of instruments that can be utilized by teams.
AI Fashions
The open-weight mannequin market is reshaping the economics of AI. This cycle introduced no less than 10 vital mannequin releases or updates throughout open and closed suppliers, with pricing strain coming from a number of instructions. DeepSeek now performs inside a fraction of Claude Opus 4.7 on coding benchmarks at a radically lower cost; Alibaba, Google, Z.ai, and Moonshot all launched succesful open fashions this cycle. The Stanford AI Index paperwork this at scale. For organizations constructing on AI, the query is not whether or not open-weight alternate options are viable however which trade-offs they’re prepared to make on value, portability, and assist.
- Google has printed an inventory of 1,302 real-world use instances for generative AI. It’s very lengthy and possibly not price studying by yourself. Nonetheless, you may need to level your agent at it.
- OpenAI has introduced GPT Photographs 2, its flagship mannequin for producing photographs. The preliminary response is that it’s barely higher than Google’s Nano Banana. What distinguishes Photographs 2 is that it “thinks” earlier than producing the picture.
- Anthropic used Claude to work on some issues in alignment analysis. Claude outperformed the people at decrease value. The issues had been, admittedly, cherry-picked to be simply scoreable. However the experiment additionally demonstrated {that a} much less succesful mannequin can supervise a stronger mannequin.
- Moonshot Labs has launched Kimi K2.6, the newest in its collection of open fashions. It additionally open sourced the Kimi Vendor Verifier, a software that checks the accuracy of distributors promoting inference utilizing Kimi.
- Alibaba has launched Qwen3.6-35B-A3B, the newest mannequin in its Qwen collection. It’s a mixture-of-experts mannequin with 3B lively parameters. Simon Willison stories that it attracts nice flamingos, when you contemplate that related.
- Anthropic has launched Claude Opus 4.7. The mannequin is positioned as an intermediate step between Opus 4.6 and Claude Mythos Preview. Anthropic claims that 4.7 is best at multimodal work, together with imaginative and prescient, instruction following, and reminiscence use. Its new tokenizer will increase the variety of tokens that Claude makes use of. As a result of billing is predicated on tokens, that’s successfully a value enhance. Simon Willison has constructed a software to check the token utilization of various fashions.
- Google has introduced Gemini 3.1 Flash TTS, a text-to-speech mannequin that offers extraordinary management over the audio system: accents, fashion, expression, and extra.
- Stanford’s 2026 AI Index Report is out, with over 400 pages of information and evaluation in regards to the state of AI.
- Meta’s refactored AI lab has launched its first mannequin, Muse Spark. It’s a multimodal mannequin that has been designed for integration with Meta’s merchandise. There’ll finally be a Considering Mode for orchestrating brokers.
- DeepSeek has launched a preview model of DeepSeek-V4, its newest open-weight mannequin. It’s a big mannequin (over 1T parameters) with efficiency very near the frontier fashions, however (as Simon Willison factors out) operating it is rather cheap.
- OpenAI launched GPT-5.5, which some are calling “Mythos-like hacking, open to all.” Along with being its “smartest and most intuitive” mannequin but, OpenAI claims that it reduces token counts, thereby decreasing value. Different sources report that, whereas it scores extremely on benchmarks, GPT-5.5 is markedly extra more likely to hallucinate and supply incorrect solutions.
- Z.ai’s GLM-5.1 is a brand new model of the open supply GLM-5 mannequin that has been optimized to carry out properly on long-running duties.
- Google has launched Gemma 4, a brand new model of its household of open supply fashions. The household features a 31B model and a mixture-of-experts model with 26B parameters, 4B lively. These are all reasoning fashions which are designed for agentic workflows. One mannequin, Gemma 4 E4B, can run on the iPhone and Android.
Software program Growth
Anthropic has clearly been profitable the announcement race. Whether or not it’s additionally profitable on efficiency is a special query. Claude Code was a favourite amongst builders till its efficiency slipped. Many switched to newly launched Cursor 3, which places an agentic interface entrance and middle whereas relegating the IDE to the background. Anthropic’s public postmortem on Claude Code’s conduct regression is price studying each for its particular findings and as a mannequin for the way AI suppliers ought to talk high quality points to builders. And Cursor’s transformation from an IDE into an agent is a sample we anticipate to see repeated throughout the business.
- OpenAI has introduced “workspace brokers.” Workspace brokers might be shared throughout a staff, whereas the brokers we have now thus far are tied to particular person productiveness. They permit a staff to collaborate on constructing shared instruments to automate workflows.
- Microsoft has introduced two new instruments, Critique and Council, that use Claude and GPT collectively to resolve analysis issues. Their benchmark outcomes present that the mix works higher than any mannequin used by itself.
- Stash is an open supply reminiscence layer that agent builders can use to attach their brokers to fashions. We’re starting to see an agentic stack that’s composed of interchangeable modules.
- Builders have been complaining a few drop in Claude Code’s conduct over the previous couple of months. Anthropic has issued a response explaining what occurred and the way they’re fixing it.
- Glif is an agent that tries to unify all of the LLMs and instruments at your disposal. You don’t should resolve which mannequin or software is finest for every activity; it makes the choice for you and will get the duty completed.
- OpenAI has decoupled its agent harness from computing and storage, enabling sturdy long-running brokers. The harness is now open supply and might be custom-made by the Brokers SDK.
- Anthropic has introduced Claude Code routines. A routine is a bundle that features a immediate, a repository, and connectors that can run robotically on Anthropic’s infrastructure, both on a schedule or when triggered.
- Anthropic additionally introduced Claude Managed Brokers, a prebuilt harness for creating brokers that run on Anthropic’s infrastructure. The harness gives a lot of the infrastructure that an agent wants (reminiscence administration, and so on.) however might be configured for the person’s duties. Anthropic’s objective seems to be changing into the AWS of agentic AI: a service supplier for software builders.
- Interoperability between instruments, fashions, and plug-ins is permitting a new programming stack to develop: an orchestration layer, an execution layer, and a evaluate layer.
- Amazon has launched an agent registry service as a part of AWS Bedrock AgentCore. Bedrock AgentCore is a set of companies that make it simple to construct and deploy brokers on AWS. The registry provides builders a technique to uncover third-party brokers that is likely to be helpful to their work.
- Bryan Cantrill’s essay on laziness is a must-read. AI isn’t lazy, and that’s an issue. When work prices nothing, there’s no want to consider future employees. Laziness is a advantage that we have to protect.
- Anthropic has introduced Claude Design, a brand new software designed to assist designers. It competes instantly with Figma and Canva. It’s at present in “analysis preview.”
- Perplexity has launched Private Pc, an area AI agent that runs on a devoted Mac mini (Home windows to return) and has persistent entry to your information, native apps, inbox, and the online.
- Anthropic has launched a Claude plug-in for Microsoft Phrase, concentrating on the authorized market. Automated edits seem as tracked adjustments.
- LiteParse is a command-line software that extracts textual content from PDF information. In the event you’ve by no means wanted to try this, you’ve lived a blessed life. Simon Willison has constructed a web-based model that runs LiteParse within the browser.
- Luke Wroblewski has stated that designers ought to code; they should perceive their medium. However round 2014, heavyweight frameworks like React and Angular acquired in the best way. Coding brokers are actually making “collapsing the hole between designing and constructing.”
- Cursor 3, the letest launch of Cursor, relegates its IDE to the background. The principle display is designed for orchestrating brokers. You possibly can fall again to the IDE for modifying code if you must.
- Within the first quarter of 2026, Apple’s app retailer has seen a enormous (84%) enhance within the variety of new apps, in comparison with the primary quarter of 2025. The trigger might be the benefit of utilizing AI to create new apps. Apple additionally seems to be limiting using “vibe coding” to create new apps, and has eliminated a number of vibe coding apps from the app retailer.
- Anthropic unintentionally leaked the supply code for Claude Code, prompting waves of commentary. Two of probably the most fascinating are Shlok Khemani’s tour of what he discovered fascinating within the supply and Gergely Orosz’s dialogue of the authorized implications.
- “The Hidden Technical Debt of Agentic Engineering” argues that, as with machine studying, brokers are comparatively small components of bigger software program techniques, and that technical debt accumulates in all of the supporting modules.
- Chat isn’t the most effective interface for working with AI. Ethan Mollick writes that the present era of AI fashions and brokers are able to creating task-specific interfaces on the fly.
Safety
Safety has spent a number of time within the information. Two core instruments for safe personal networking, Tor and Sign, have been attacked. In each instances, the assault didn’t contain the software program or protocols themselves. These assaults train us that safe techniques are sometimes jeopardized by the software program that surrounds them. We’ve additionally seen that ransomware gangs are utilizing postquantum encryption, and that quantum computer systems are more likely to break conventional encryption ahead of anticipated. In the event you’re not investing in safety, it’s time to start out.
- The Tor community is the gold customary for safe personal networking. Researchers not too long ago found a vulnerability in Firefox browsers that lets attackers de-anonymize identities. The vulnerability has been fastened in Firefox 150, however it’s a reminder that something might be attacked.
- Everyone knows that ransomware gangs use encryption. The Kyber group is making the transition to postquantum encryption.
- A provide chain assault towards npm permits unhealthy actors to steal builders’ credentials. As soon as it has contaminated a sufferer, it inserts itself into different packages that the sufferer publishes.
- Legislation enforcement businesses had been briefly in a position to exploit a vulnerability in iOS notifications that allowed them to entry unencrypted messages despatched with the Sign safe messaging system. The vulnerability has been patched. It’s essential to know that the vulnerability wasn’t in Sign itself however within the setting by which it operated.
- With AI, time from discovery of a vulnerability to exploitation has dropped to zero. To assist protection catch up, Google has added three brokers to its Google Safety Operations platform: Menace Searching, Detection Engineering, and Third Social gathering Context.
- Microsoft stories that criminals are more and more utilizing Groups to impersonate assist desk personnel, who ask customers for his or her credentials after which steal information.
- NIST has stopped assigning severity scores to lower-priority vulnerabilities. All vulnerabilities will nonetheless be added to the Nationwide Vulnerability Database (NVD).
- The NSA is utilizing Claude Mythos Preview, regardless of Anthropic being blacklisted by the Pentagon. Anybody need to guess what they’re utilizing it for?
- Anthropic will ask for identification verification in some instances.
- Small open-weight fashions can do in addition to Anthropic’s Mythos at discovering vulnerabilities. The important thing isn’t the mannequin; it’s the system inside which the mannequin works.
- A new malware marketing campaign embeds credit-card stealing software program right into a single pixel SVG picture. ecommerce websites utilizing Magento Open Supply or Adobe Commerce are susceptible.
- Anthropic has pulled its latest mannequin, Claude Mythos, from broader launch as a result of it’s too good at discovering vulnerabilities in different software program. They’ve made it obtainable to just a few companies by way of Challenge Glasswing, an try to safe vital software program earlier than it may be exploited. The AI Safety Institute’s evaluation of Claude Mythos Preview says that it “represents a step up over earlier frontier fashions in a panorama the place cyber efficiency was already quickly enhancing.”
- Many open supply safety maintainers agree with Greg Kroah-Hartmann‘s report that the standard of AI-generated safety bug stories has gone up tremendously.
- Variations of Claude Code that embody the Vidar malware have been printed on GitHub. They’re primarily based on the code that Anthropic inadvertently leaked. These variations entice victims to obtain them by claiming to have unlocked enterprise options.
- Claude has been used to uncover zero-day distant code execution vulnerabilities in each Vim and Emacs. The vulnerabilities are triggered when a person opens a file. An replace is obtainable for Vim; Emacs builders argue that it’s actually a bug in Git, which can be appropriate however misses the purpose.
- Breakthroughs in quantum computing imply that computer systems able to cracking present encryption algorithms could also be on the horizon.
Infrastructure and Operations
A number of suppliers launched overlapping items of an agent stack this cycle, protecting orchestration, persistence, reminiscence, and registry companies. A 3-layer mannequin (orchestration, execution, evaluate) is changing into the usual structure, however every vendor’s implementation makes totally different bets about portability and sturdiness. It’s essential to guage every vendor’s merchandise fastidiously earlier than deciding on an agent stack.
- Microsoft now permits admins to uninstall Copilot, although there are circumstances.
- Google has introduced two new eighth-generation TPUs. One is designed for coaching (8t), the opposite makes a speciality of inference (8i). That is the primary time Google has produced specialised TPUs for coaching and inference.
- Google has open-sourced Scion, its testbed for agent orchestration.
- Anthropic has agreed to purchase 3.5 gigawatts of computing energy from Google and Broadcom, maker of Google’s GPUs. The deal specifies energy consumption quite than the variety of chips, implying that the limiting issue isn’t computation however the availability of energy. Chips come and go; watts are a continuing.
- Ollama now makes use of Apple’s MLX framework to enhance efficiency on Apple silicon. Assist is at present restricted to the Qwen3.5-35B-A3B; assist will probably be added for different fashions. As a part of this replace, it additionally makes use of NVIDIA’s NVFP4 floating level format for mannequin quantization.
Net
Don’t overlook the online layer when planning for AI-driven disruption. The online’s infrastructure is older than most people who keep it, and a number of other objects this cycle are reminders of the hole between what that infrastructure was designed for and the way it’s used at this time. Two take care of protocols which have outlasted their unique assumptions; one other reimagines the dominant CMS from scratch utilizing present tooling.
- Is PHP the brand new COBOL? What about open supply itself? “Who Will Keep the Net When PHP’s Veterans Retire?” factors to a actuality that we don’t like to consider. Not solely are firms reluctant to rent junior builders; those they do rent aren’t studying older applied sciences.
- Laravel is outwardly injecting advertisements for its business cloud service into brokers. What occurs when an open supply framework receives enterprise funding and begins injecting advertisements into brokers? We’re about to search out out.
- Doesn’t each musician want instruments to typeset Gregorian chant?
- Is IPv8 the way forward for the Web? IPv6 has been “two years away” since early within the Nineties. IPv8 is absolutely backward suitable with IPv4, and resolves its safety and deal with depletion points.
- Cloudflare has launched EmDash, an alternative choice to WordPress primarily based on how the online is used at this time. Drew Breunig calls this a reimagining: a brand new part of software program improvement by which we are able to use agentic programming to rethink and reimplement instruments primarily based on present wants.
- Is BGP Secure But? is an online app that checks whether or not your ISP has applied BGP (the protocol that’s accountable for routing packets at web scale) appropriately. Many haven’t.
Biology
- OpenAI has introduced GPT-Rosalind, a mannequin that has been tuned for 50 frequent workflows in biology. In contrast to most fashions, Rosalind has been tuned to be skeptical quite than enthusiastic or sycophantic. Entry to Rosalind is restricted due to the potential for hurt.
Robotics
- Spot, the Boston Robotics robotic canine, can now learn gauges and thermometers. It makes use of the Gemini Robotics-ER 1.6 mannequin, which might motive about visible info.
- Main League Baseball is utilizing a robotic system to rule on challenges to a human umpire’s ball/strike calls.
