The US Secret Service (USSS) dismantled a network of over 300 co-located SIM servers and 100,000 SIM cards positioned within 35 miles of the UN General Assembly in New York that were used to conduct telecommunications-related threats against senior US government officials. The devices were capable of performing multiple attack vectors, including disabling cell phone towers, facilitating denial of service attacks, and enabling anonymous encrypted communications between nation-state threat actors and criminal enterprises. Early forensic analysis revealed cellular communications between nation-state threat actors and individuals known to federal law enforcement, representing an imminent threat to protective operations during the high-profile international gathering.
A researcher at Carnegie Mellon University demonstrated that an LLM can coordinate a system of agents to recreate real-world attacks. As part of his PhD, Brian Singer recreated the network environment of the 2017 Equifax data breach and observed that an LLM could replicate the attack. These capabilities allow smaller organizations to use LLMs as dedicated red teamers.
Microsoft has removed a compatibility hold that prevented devices with integrated cameras from installing Windows 11 24H2 due to a face detection bug that caused app freezes when using the Camera app, Windows Hello facial recognition, and other camera-utilizing applications.
MI6 launched a dark web portal called Silent Courier to allow potential agents worldwide, particularly Russians, to securely share sensitive information about terrorism and hostile intelligence activities.
Steam removed the free-to-play game BlockBlasters after malware stole over $150,000 in cryptocurrency from players, including $32,000 from a Latvian streamer with stage 4 cancer raising funds.
Sponsored
You can't afford to ignore LLMs. Your SEO playbook is still relevant - but it needs a version update, pronto.
EU lawmakers are considering changes to the e-Privacy Directive that triggered the constant cookie consent pop-ups online. Ideas include browser-based consent settings or broader exceptions for essential cookies, reducing the need for repeated prompts. Industry leaders are pushing for alignment with GDPR's risk-based model, while privacy groups raise concerns that data protections could be weakened.
ChatGPT drives high-converting traffic but often generates faulty URLs, which can waste potential visits. An Ahrefs study found ChatGPT 5 links to error pages nearly three times more than Google Search. To recover lost traffic, start by tracking ChatGPT-driven 404 pages in Google Analytics using filters for session source and error page titles. Next, create helpful 404 pages that guide visitors when they land on hallucinated URLs. Use 301 redirects only for broken URLs that generate measurable traffic.
A custom social feed was built by outlining a framework with features like voting, saves, topic filters, timestamps, breaking news alerts, and weekly reading stats. Lovable handled the technical build with Supabase, PWA support, and responsive design, while conversational feedback shaped iterations and aesthetics. Specific news and culture sources were added, unwanted ones excluded, and a widget was placed on the phone's homepage. The result is a curated, distraction-free content hub distinct from traditional social media.
Sponsored
Most marketing teams were early adopters of ChatGPT and Claude, using them for campaign briefs, social copy and more. But AI can do more than content generation.
Go Fish Digital is hosting a webinar that will cover how brands can get cited by top publications to boost brand visibility, SEO, and inclusion in LLM responses. The panel will discuss how PR drives visibility, the links and mentions that carry weight with LLMs, and how to influence knowledge panels and entity-driven results. The event takes place today at 2 pm ET.
Tapestry is an app that gathers content from blogs, RSS feeds, YouTube channels, and other sources into a single chronological timeline. No algorithm interferes with the feed. It prioritizes privacy with no tracking, allows muting of topics and spoilers across feeds, and offers customization of layouts, fonts, and icons.
LinkedIn's Company Intelligence API shifts B2B measurement from lead-level tracking to company-level insights that reflect how real buying decisions are made. Early adopters reported a 287% increase in companies reached, 75% more MQLs, 96% more SQLs, and a 43% reduction in acquisition costs. The tooling allows marketers to see the full buyer journey, prioritize high-value accounts, improve attribution accuracy, and link campaigns directly to pipeline growth and revenue.
Brands often fall into the “Tyranny of the OR,” choosing single-dimensional strategies rather than embracing paradoxical promises that deliver both seemingly conflicting benefits. Unsuccessful examples include Cracker Barrel returning to its original premise and Jaguar abandoning its legacy for a new identity, each potentially missing opportunities to balance heritage with innovation. Research shows consumers dislike trade-offs that force them to sacrifice something they value, preferring brands that satisfy multiple, even contradictory, needs at once. Successful brands like McDonald's, Apple, and Coke have grown by solving paradoxes rather than choosing one side.
This study examined the cognitive impact of using ChatGPT for essay writing. It compared participants who used an LLM, a search engine, or only their own brain across multiple sessions. EEG data revealed that reliance on the LLM reduced neural connectivity and engagement compared with the Brain-only and Search Engine groups, and participants using the LLM showed weaker memory recall and lower ownership of their work. While LLMs offer immediate assistance, extended use may impair cognitive effort and learning skills.
The head of content, social, and integrated marketing at Microsoft shares what's worked for their EGC program.
Examples of gamified emails built via Zaymo.
The new beta for Google Ads lets app marketers apply seasonality signals to Smart Bidding, enabling proactive bid boosts during major promotions without waiting for algorithm changes.
With core context engineering principles and frequent intentional compaction, current AI models can be great at handling large, detailed codebases. AI can rarely produce correct code in one shot in large code bases, so it needs steps before implementation to do research on the codebase and plan its steps.
The "thundering herd" problem occurs when a cache miss leads to a large number of concurrent requests hitting the database for the same record, overwhelming it. There are two solutions: using a distributed lock to allow only one request to access the database, or using in-process synchronization with (as an example) Java's `CompletableFuture` and `ConcurrentHashMap`. While the distributed lock works across multiple nodes, the in-process synchronization avoids network calls but doesn't coordinate across nodes.
2025.09.24·distributed-computing-musings.com·8 min read·TLDR Web Dev Infobip's infrastructure team handles 10 billion messages daily. They overcame issues like email overloads, data center failures, and scaling for WhatsApp Business by using cloud solutions, re-architecting systems, and adopting Kubernetes. A critical decommissioning mistake in 2024 led to the implementation of delayed destructive actions, preventing data loss.
While appreciating HTMX's potential to reduce JavaScript reliance, this dev found its lack of structure problematic, leading him to create MESH, a modular SSR framework based on the "one component = one endpoint" principle. He concludes that he's not using HTMX in the way it's intended and drops it in favor of simpler JavaScript and SSE.
Serving appropriately sized images for different screen sizes avoids wasting bandwidth and compute power. This can be done by using the `srcset` and `sizes` attributes on the `img` tag to specify multiple image sources and their intended widths based on viewport size.
Devs should test their websites without JavaScript enabled to check accessibility and resilience. JavaScript can fail due to slow networks, browser extensions, or user settings, so sites should use progressive enhancement, starting with semantic HTML that works without JS and adding JavaScript features as enhancements.
Sponsored
From token storage to session handling, this guide walks through the security practices every frontend developer should know. Learn which approaches protect your users and which mistakes to avoid when building authentication into modern web applications.
TanStack Start has reached its v1.0 Release Candidate. This version includes type-safe routing, server-first functions, built-in streaming, and URL-as-state primitives, all while offering flexibility for both SPA and SSR architectures.
The AI Data Generator helps users generate sample data with various input parameters, export it in various formats, and explore it within Metabase.
Active Agent is a Ruby gem that helps integrate AI features into Rails applications using familiar Rails conventions and patterns. It introduces "agents" as a new abstraction that encapsulates AI-backed logic, using action-driven objects, callbacks, and prompt rendering similar to Rails controllers and mailers.
PostgreSQL's original design goals from 1986 have shaped its modern success. These goals included support for complex objects, extensibility, active database features, and crash recovery, laying a flexible framework. Postgres' adherence to these principles has cemented its position as a versatile and trusted database today.
VCs are aggressively pursuing investments in top AI startups with gifts and preemptive funding rounds. Companies like Decagon AI are fielding unsolicited offers at soaring valuations. This frenzy is fueled by the belief that AI startups can disrupt tech giants, with a large portion of venture capital flowing to a small group of frontrunners.
Manus is an AI agent that had improved performance through context engineering techniques. The devs optimized the KV-cache hit rate by maintaining a stable prompt prefix with append-only context. They advise against dynamically changing the action space. The file system should be used as an extended memory, along with recitative methods to manipulate attention.
CSS can be used to target specific characters for styling through manipulating fonts and Unicode ranges, though the practical applications are limited.
AI coding tools like Gemini CLI are being used for software development by automating code generation from natural language requirements and shifting the developer's role towards architecture and problem decomposition.
A recent NPM attack, the Qix incident, shows the risk of malicious code injected into widely-used packages targeting end-users by way of bundled website code.
Redux is still a reliable choice for complex React projects due to its predictability, observability, and architectural clarity.
Sponsored
They say ignorance is bliss – until it's preventing you from expanding your app to non-English speaking markets. Then you find yourself struggling with tools that fail to understand tone, misuse jargon, and ultimately lose your finely honed brand voice in translation.
Oracle will oversee the security of Americans' data and monitor changes and updates to TikTok's recommendation technology under a new deal. A copy of the recommendation engine that powers the app will be licensed from China to an American investor group that will oversee the app in the US. Oracle, along with private equity firm Silver Lake, will also invest in the new American TikTok. The US-run TikTok will work to retrain the algorithm on users' data in the US. China will not have access to the data.
2025.09.23·links.tldrnewsletter.com·3 min read·TLDR Nvidia and OpenAI have a new partnership that includes plans for an enormous data center buildout and a $100 billion investment by Nvidia into OpenAI. The deal will allow OpenAI to build and deploy at least 10 gigawatts of Nvidia systems for its artificial intelligence data centers to train and run its next generation of models. It is a bet on continued model improvements that runs contrary to the view of some skeptics, who say model development is hitting a wall. Nvidia will make its investment in OpenAI progressively as each gigawatt is deployed.
2025.09.23·links.tldrnewsletter.com·6 min read·TLDR Blue Origin's Blue Alchemist is designed to transform lunar and Martian regolith into solar power systems, breathable oxygen, propellant, metals, and construction materials. Ensuring sustainability requires that missions rely on local resources to provide basic necessities. Blue Alchemist uses a molten regolith electrolysis reactor, which uses electrical current to separate oxygen from metals without water, toxic chemicals, or carbon emissions. Blue Origin is on track to scale the system to make lunar landings up to 60% cheaper and reduce fuel-cell and battery masses by up to 70%.
2025.09.23·spectrum.ieee.org·2 min read·TLDR Karuna Therapeutics and Bristol Myers Squibb's Cobenfy pill is the first novel medication for schizophrenia in 50 years. The drug, approved by the FDA last fall, reduces symptoms like hallucinations without the adverse effects common in other antipsychotics, such as weight gain. Cobenfy's approval was the culmination of an effort that began over 15 years ago. Karuna Therapeutics and Bristol Myers Squibb Won the 2025 Gizmodo Science Fair for their research.
2025.09.23·gizmodo.com·5 min read·TLDR Sponsored
Join Snyk on October 22 [9] for this one-day event to hear from leading AI and security experts. The agenda includes keynotes, a hands-on AI Demos track, sessions on safeguarding apps and models, and
OpenAI's Responses API unlocks persistent reasoning, hosted tools, and multimodal workflows for GPT-5. It is tailor-made for reasoning models and the agentic future. While Chat Completions isn't going away, OpenAI expects Responses to eventually become the default way developers build with OpenAI models.
2025.09.23·developers.openai.com·6 min read·TLDR Very few apps get offline support right because syncing is hard. Building a local-first app means effectively creating a distributed system where multiple devices can mutate data independently, sometimes offline, while still converging to the exact same state without losing data. This involves solving the challenges of unreliable ordering and conflicts. SQLite is a battle-tested, lightweight, and readily available solution for local-first apps that guarantees convergence across devices, regardless of the sync order.
2025.09.23·marcobambini.substack.com·7 min read·TLDR Cloudflare launched in 2010 when less than 20% of the Internet was encrypted - today, well over 95% is encrypted, thanks in part to its efforts. The basic business model of the Internet for the past 15 years has been the same: create compelling content, find a way to be discovered, then generate value from the resulting traffic. There's always been a reward system that transferred value from consumers to creators. Ads were previously the only micropayment system that worked at scale to encourage an open Internet while also compensating content creators for their work, but change is being driven by AI. This article looks at how Cloudflare plans to navigate this change and what the new Internet business model will look like.
2025.09.23·blog.cloudflare.com·9 min read·TLDR