Search the archive

you one less thing to lose sleep over. Try Bitwarden free today

KUBERNETES REMOTE CODE EXECUTION VIA NODES/PROXY GET PERMISSION

An authorization bypass in Kubernetes allows service accounts with nodes/proxy GET permissions to execute commands in any Pod across the cluster, potentially leading to full cluster compromise. The vulnerability stems from the Kubelet authorizing WebSocket connections based on the initial HTTP GET handshake rather than verifying CREATE permissions for the /exec endpoint, affecting 69 Helm charts, including Prometheus, Datadog, Grafana, and Cilium. Kubernetes Security Team closed this as "Won't Fix (Working as Intended)," recommending migration to KEP-2862's fine-grained authorization when it reaches GA in April.

NETSUPPORT MANAGER 0-DAY VULNERABILITIES ENABLE REMOTE CODE EXECUTION

Two critical vulnerabilities in NetSupport Manager (CVE-2025-34164 and CVE-2025-34165) can be chained to achieve unauthenticated remote code execution through the software's undocumented broadcast feature, which operates on TCP port 5405 without requiring authentication. The flaws—a heap-based out-of-bounds write and a stack-based out-of-bounds read—enable attackers to bypass ASLR, perform arbitrary memory writes, and gain remote shell access, posing a significant risk to Operational Technology (OT) environments where the software is commonly deployed. Organizations should upgrade to version 14.12.0000 or later and restrict access to port 5405 as an interim mitigation.

CANVA AMONG ~100 TARGETS OF SHINYHUNTERS OKTA IDENTITY-THEFT CAMPAIGN

ShinyHunters is running an Okta single sign-on credential-stealing campaign against roughly 100 high-value organizations, including Canva, Atlassian, RingCentral, and others. The group uses evolved voice phishing to capture SSO logins, enroll its own devices in MFA, pivot into SaaS apps, exfiltrate data, and then extort victims. There is no confirmation of which named firms were successfully breached.

WHO OPERATES THE BADBOX 2.0 BOTNET?

A leaked screenshot from Kimwolf botnet operators revealed they had compromised the Badbox 2.0 control panel, exposing seven authorized user email addresses that OSINT investigation traced to Chen Daihai and Zhu Zhiyu of Beijing Astrolink Wireless Digital Technology—individuals whose domains were previously flagged in HUMAN Security's Badbox 2.0 report. The investigation used password reuse across breach databases, domain registration records, and social media pivots to connect qq.com email addresses to multiple Chinese technology companies distributing pre-infected Android TV boxes. This unauthorized access allows Kimwolf to bypass residential proxy mitigations by loading malware directly onto Badbox 2.0's 10+ million compromised devices.

HOW TO ENCRYPT YOUR PC'S DISK WITHOUT GIVING THE KEYS TO MICROSOFT

BitLocker encrypts Windows PCs to protect data, but many systems automatically store recovery keys with Microsoft when users sign in with a Microsoft account, creating a potential privacy risk if authorities obtain those keys. By upgrading to Windows 11 Pro, users can fully control BitLocker, decrypt any existing Microsoft-managed setup, and re-encrypt the drive while saving the recovery key locally, such as on paper or an external drive, rather than in the cloud.

LLM KEY SERVER: PROVIDING SECURE AND CONVENIENT ACCESS TO INTERNAL LLM APIS

Mercari's AI Security team developed a key server to address the challenge of managing API key-based access to LLMs. The key server uses Google Workspace and Google Cloud for OIDC authentication to LiteLLM, which provides time-limited API access to various models through a unified API. The team also developed an internal CLI, GitHub Action, and Google Apps Script template to support adoption of the LLM key server.

AGENT OS

Agent OS applies operating system concepts to AI agent governance, providing kernel-level policy enforcement that intercepts and blocks actions before execution rather than relying on prompt-based safety. The framework includes POSIX-inspired primitives such as signals, a virtual filesystem for agent memory, cross-model verification, inter-agent trust protocols with cryptographic signing, and integrations with LangChain, CrewAI, and OpenAI Assistants. An MCP server enables integration between Claude Desktop and tools for verification and kernel execution.

AWESOME HACKING

Awesome hacking is a curated list of hacking tools for hackers, pentesters, and security researchers. Its goal is to collect, classify, and make awesome tools easy to find by humans, creating a toolset you can checkout and update with one command.

ZEROPULSE

ZeroPulse is a comprehensive C2 platform that utilizes CloudFlare Tunnel technology. It is designed for secure remote management and monitoring.

WHY HAS MICROSOFT BEEN ROUTING EXAMPLE.COM TRAFFIC TO A COMPANY IN JAPAN?

Microsoft's autodiscover service was found routing email traffic for example.com—a domain reserved for testing under RFC2606—to subdomains belonging to Sumitomo Electric in Japan, potentially causing test credentials to be sent outside Microsoft's network. The misconfiguration, which may have persisted for five years, has been suppressed, but Microsoft has not explained how the Japanese company's servers were added to its network configuration. The incident raises concerns about other potential misconfigurations following Microsoft's 2024 breach, where forgotten test account privileges enabled Russian state hackers to monitor executive emails.

BROWSER-BASED ATTACKS HIT 95% OF ENTERPRISES — AND TRADITIONAL SECURITY TOOLS NEVER SAW THEM COMING

95% of organizations experienced browser-based attacks last year that evaded traditional security tools. Attackers operated inside trusted sessions where web gateways, cloud access brokers, and endpoint protection lose visibility after login. Recent incidents, including ShadyPanda's weaponized extensions and Cyberhaven's supply chain compromise, demonstrate how attackers exploit browser auto-updates and session tokens rather than zero-days. Security leaders recommend inventorying extensions, implementing 48-72 hour auto-update delays, and moving data protection to the browser layer, where 64% of encrypted traffic currently goes uninspected.

NOYB WIN: MICROSOFT ORDERED TO STOP TRACKING SCHOOL CHILDREN

Austria's data protection authority ruled that Microsoft illegally placed tracking cookies on a pupil's device through Microsoft 365 Education, using them to analyse behaviour, collect browser data, and support advertising. The authority ordered Microsoft to stop tracking the child within four weeks and rejected Microsoft's attempt to shift responsibility to its Irish subsidiary.

HACKERS CAN BYPASS NPM'S SHAI-HULUD DEFENSES VIA GIT DEPENDENCIES

Researchers discovered "PackageGate" vulnerabilities in JavaScript package managers that allow bypassing npm's '--ignore-scripts' defense via malicious .npmrc files in Git dependencies, with pnpm, vlt, and Bun issuing patches while npm rejected the report as expected behavior.

GOOGLE WILL SETTLE ITS ASSISTANT SPYING LAWSUIT FOR $68 MILLION

Google has agreed to pay $68 million to resolve a class action lawsuit over “false accepts,” when Google Assistant devices recorded audio without users intentionally saying “Ok Google.”

'STANLEY' TOOLKIT TURNS CHROME INTO UNDETECTABLE PHISHING VECTOR

A new malware-as-a-service kit called Stanley generates malicious Chrome extensions that overlay convincing phishing pages on top of real sites while keeping the legitimate URL visible through a fake productivity extension.

OXYGEN MONOPOLIES

Fast news cycles and rapid AI adoption mean startups win by owning a clear category narrative early, since attention spikes fast and hardens just as quickly. AI companies now reach $10M+ ARR in under a year, and some hit $100M ARR with fewer than 100 employees, which makes early brand positioning a real competitive advantage. Strong brand association lets one company suck up most of the oxygen, as seen when it becomes shorthand for its category, like Harvey for legal AI. Brand recall now matters more than awareness, since even Tide holds 41% market share but ranks 9th on Amazon search. Companies must make people think of it first or risk being forgotten.

INTRODUCING INVISIBLE ENGAGEMENT AND WHY IT MATTERS

Most consumer interaction with brand content happens through invisible actions that platforms do not track, which makes visible metrics misleading. In a study of 12,855 social media posts, only 0.6% showed visible engagement, while 22.3% showed invisible engagement and 77.1% showed none. The most common invisible actions were ‘see' at 42% and ‘stop' at 41%, which signal attention without public interaction. Posts linked to invisible actions such as ‘see', ‘search', or ‘save' had a higher median shop engagement rate of 10.1% compared to 6.8% for posts that received a ‘like'. These findings show that invisible engagement is more closely tied to purchase behavior and should play a larger role in how marketers measure performance.

4 CTA BUTTON TESTS

Focus on first-person, personalized CTAs that reflect the user's role, interest, or goal, for example, “I want top HR tips” or “Save 25% on my wellness journey”. Use playful or contrasting button pairings, like “I hate learning new things” instead of “Save my seat”. Avoid generic verbs such as “Download” or “Submit”. Replace them with engaging, action-oriented phrases like “Get my discount” or “Let's go!” The goal is to make the CTA feel personal and clearly tied to user benefit.

ASK USERS TO REGISTER AT THE START (NOT THE END)

Asking users to register at the start of shopping increases registrations, purchases, and long-term revenue. In a field experiment of over 75,000 users, early registration made shoppers 58.1% more likely to register and 10.9% more likely to buy. Those shoppers bought 16.8% more items and generated 13.2% higher total revenue over a year. Casual shoppers browsed 12.3% less after early registration prompts. Optional registration prevented short-term sales losses.

52 GTM SIGNALS

A screenshot that lists out 52 GTM triggers across 4 datasets. Start by segmenting your ICP and understanding their buyer journey. Use triggers from company activity, news, job postings, or tech changes to time outreach. Focus on high-quality, relevant data. Each signal should guide what to watch, why it matters, and how to craft your message. Right-click the screenshot and select “Open in new tab” to see a zoomed-in view.

12 STEPS TO WINNING BRAND MENTIONS IN LLMS

This guide from Noble outlines 12 steps to winning LLM visibility. Start by identifying the questions your brand should appear for. Check what LLMs currently say and which sources they cite. Expand your list with high-ranking, relevant pages from Google. Clean and score all URLs to prioritize targets. Build a contact list for authors or content owners. Decide the type of mention or update needed. Craft personalized outreach at scale. Manage replies and negotiate terms or swaps. Provide easy-to-use content snippets. Implement any reciprocal mentions on your site. Verify that mentions go live and impact LLMs.

HOW DO YOU COMPETE IN AGENTIC COMMERCE?

Agentic commerce is reshaking e-commerce by turning organic search into a verification gate for AI-driven purchases rather than a traffic channel. AI agents powered about 20% of retail sales during the 2025 holiday season, and major platforms now support direct checkout through commerce protocols that bypass traditional websites. Autonomous buying mainly works in a narrow-middle since high-priced purchases resist delegation and low-priced goods already rely on subscriptions, so the real value comes from conversational commerce that compresses an average 14-click journey into 1 or 2 steps. The protocols force merchants to optimize structured product data and APIs instead of pages. As a result, SEO shifts from driving clicks to feeding trusted, granular data into AI systems, favoring incumbents for broad intent and specialists for specific intent.

THERE ARE REASONS TO BE BULLISH ABOUT AI CHAT ADS

AI chat ads could reshape retail media, but their success is uncertain. OpenAI and Google are testing ads in ChatGPT and Gemini, while Amazon and Walmart explore AI-powered shopping. Experts say AI ads could combine strong intent and deep user knowledge, offering personalized experiences if they feel helpful rather than pushy. Retailers want to stay the merchant of record and control logistics, limiting AI's transactional role. Widespread adoption could disrupt retail media by capturing intent and transaction data, but user experience and partner balance remain key challenges.

WHY SEARCH AND SHOPPING ADS STOP SCALING WITHOUT DEMAND

Search and Shopping ads capture existing demand but don't create it. Growth stalls when the audience or queries are limited, not because campaigns are broken. Broad match and AI Max can stretch coverage slightly, but they still rely on search activity. True growth requires creating demand upstream through content, SEO, social, and paid media.

PLEASE STOP SELLING LIKE IT'S 2021

In 2026, B2B wins with self-serve, zero-friction value, or buyers will build it themselves with AI.

5 EMAIL PREDICTIONS FOR THE CHARITABLE SECTOR

Nonprofits should show gratitude year-round, focus on story-driven content, and ditch generic mail-merged templates.

META IS PREPARING TO TEST PREMIUM SUBSCRIPTIONS ACROSS ITS APPS

The strategy targets power users and everyday consumers alike, betting that meaningful control and AI features can justify payment despite growing subscription fatigue.

APPLE'S AIRTAG 2 IS EASIER TO FIND THANKS TO NEW CHIP

Apple's new AirTag, simply dubbed 'the new AirTag', features a new Bluetooth chip that offers substantial improvements. It can work with the Precision Finding feature in the Find My app to find items up to 50% farther away. The speaker is also 50% louder. The new AirTag doesn't introduce any major new features, only updates to the device's accuracy, volume, and range.

MICROSOFT INTRODUCES NEWEST IN-HOUSE AI CHIP

The Microsoft Azure Maia 200 is a new in-house AI server chip designed for inferencing AI models with ludicrous speeds. It is the most efficient inference system Microsoft has ever deployed. The chip gives 30% more performance per dollar and 50% higher TDP than the first-gen Maia 100. It can hit up to 10 petaflops of FP4 compute, three times higher than Amazon's Trainium3.

BUILDING BRAINS ON A COMPUTER

Before we simulate human brains, we first need to demonstrate mastery at the sub-million-neuron-brain level. Achieving this goal will depend primarily on more and better data, which is bottlenecked by funding and the achievement of some key technological breakthroughs. It will soon be feasible thanks to a couple of recent advances. We are nearly capable of overcoming such bottlenecks at the scale of insect-sized organisms.

SEEING LIKE A SEDAN

Tesla robotaxis navigate using an array of cameras, while Waymos uses LIDAR technology along with a suite of complementary sensors. Tesla challenged convention, but it has quietly reintroduced radar, and it is possible it may bring in other modalities. The paths for the two companies seem likely to converge in the future. This means that instead of debating whether cameras or LIDAR is better, regulators should be asking whether robotaxis in general are as safe as human drivers.

levels. Download your copy

A FEW RANDOM NOTES FROM CLAUDE CODING QUITE A BIT LAST FEW WEEKS

AI agent capabilities crossed a threshold at the end of last year. This has caused a phase shift in software engineering. The intelligence part is a bit ahead of integration, so there is a need for new organizational workflows and processes. This year will be a high-energy year as the industry digests the new capability.

37SIGNALS ISN'T SMARTER THAN YOU, BUT THEY ARE DIFFERENT

37signals' engineering strategy results in good code. The company stays small, cuts scope ruthlessly, and hires well. Its strategy creates a flywheel where developers maintain less code, allowing them to apply more fixup/review cycles, creating a cleaner code base, which helps to reduce headcount. Most companies can't adopt the strategy because they think more features equals more money.

THE ADOLESCENCE OF TECHNOLOGY

AI models are now beginning to make progress in solving unsolved mathematical problems. It is so good that strong engineers are now handing their coding to AI. Similar improvements are occurring across biological science, finance, physics, and a variety of agentic tasks. If progress continues, it will only be a few years before AI is better than humans at essentially everything. This could result in a serious civilizational challenge that will require decisive and careful action.

THE BEHAVIORAL COST OF PERSONALIZED PRICING

Today's personalized pricing is different. The average person probably doesn't even know that haggling is an option or the factors that contribute to personalized pricing. In the future, people will cultivate their digital representations so they can get the lowest prices and best treatment. This access could even be sold to others who aren't so careful.

CLOSING THE SOFTWARE LOOP

The Holy Grail of software development is full self-improvement, and the key to building this is a system that can autonomously generate bug reports and understand features that users will want.

PERFORMANCE HINTS

This post is a review of a blog post written by Jeff Dean and Sanjay Ghemawat, the power duo who arguably made Google what it is today.

SPACEX STARSHIP V3 GETS LAUNCH DATE UPDATE FROM ELON MUSK

Starship Flight 12 will launch in six weeks.

SAAS, WIDELY MISUNDERSTOOD

Replacing SaaS with AI would cost more and deliver a lower-quality product in almost all scenarios, so it makes no sense from an economic or product perspective.

TSMC RISK

The only way to actually meet demand in 2028 and beyond is to have competition in the foundry space.

LIFE ON CLAUDE NINE

Claude makes problems that used to take hours take minutes, automating every tedious task away.

spot →

Can't join live? Register anyway [4], and we'll send you the recording so you can watch on your own time.

ANTHROPIC INTEGRATES INTERACTIVE MCP APPS INTO CLAUDE

Anthropic has introduced MCP app integration into Claude's AI assistant, enabling seamless use of tools like Asana, Slack, Figma, and Box within its interface. This update enhances real-time interaction, setting Claude apart by embedding live app experiences directly into chats, boosting workflow efficiency. The move supports open standards and encourages developer-created integrations, aligning with Anthropic's ecosystem goals.

QWEN3-MAX-THINKING DEBUTS WITH FOCUS ON HARD MATH, CODE

Qwen3-Max-Thinking is a reasoning model for complex math, coding, and multi-step workflows. Available on Alibaba Cloud's Model Studio, this model excels in tasks that require evidence gathering and deep verification, offering adaptive tool-use and built-in web search features. Early testers highlight its benefits for developers and enterprises needing long-context reasoning and tool-using agents.