GitHub Agentic Workflows enable repository automation by running AI coding agents within GitHub Actions. Users define these continuous AI workflows in natural language using Markdown files, which are then compiled into GitHub Actions to automate tasks like code improvements, issue triage, documentation, and analysis. They integrate deeply with GitHub, support multiple AI engines like Copilot, Claude, and Codex, and prioritize safety with sandboxed execution and minimal permissions.
Matchlock is a CLI tool designed to run AI agents within secure, ephemeral microVMs, preventing unrestricted access to your machine. It achieves this by providing network allowlisting, injecting secrets in-flight (so they never enter the VM), and blocking all other connections by default. Agents operate in a full, isolated Linux environment that boots quickly and is disposable, ensuring that even malicious actions don't compromise your credentials or data.
Apple silicon Macs achieve their high performance mainly due to their Efficiency (E) cores, which handle numerous background tasks, rather than solely relying on Performance (P) cores. Immediately after startup, E cores are heavily used for processes like Spotlight indexing and initial backups, leaving the P cores largely free for user applications. This architecture is managed by macOS's Quality of Service (QoS) system, which prioritizes foreground tasks for P cores and restricts background tasks to E cores, even when P cores are idle.
Cloud agents like Devin shift AI coding from local pair programming to organizational delegation. Running on remote infrastructure, these tools allow non-engineers to trigger tasks via Slack or Jira. This model scales capacity asynchronously, moving the primary bottleneck from code generation to automated review.
This dev tested Claude on three real React/Next.js bugs ranging from easy to hard. It aced the simple one, half-fixed the medium one (but introduced new issues), and completely failed the tricky one, confidently going in circles with wrong fixes. AI is a great first pass for standard bugs, but knowing when to stop prompting and start thinking is the real skill.
AI's ability to produce merely "good enough" software (or "slop") might create widespread user apathy, ultimately leading to the demise of quality software craftsmanship.
Vertical software stocks are down 43% from their highs. DevTools are down 21%. The gap doesn't track with which sectors AI can actually replace - it correlates almost perfectly with growth rates. Slow growers (8-11%) are getting crushed while faster growers (21-22%) are holding up. The market isn't pricing in AI disruption. It's pricing in who can outrun it.
AI companies are fighting churn and facing a retention apocalypse. The real retention curve starts when the hobbyists leave, leaving only the people who've integrated AI into actual work. AI churn is higher because it's easy to buy and cancel, and many trials reflect curiosity more than long-term commitment. Companies that win move upmarket, build for habitual workflows, and add services/FDEs on top of AI.
ZoomInfo was built from $0 to IPO, and its founder has a nuanced take on the "when do you upgrade your early team" question. His answer: every exec was a first-timer once, and external hires aren't automatically better. The real question is whether someone can do what the business needs for the next 12-24 months, not whether they've done it before.
Modern systems optimize by simulating terminal states, collapsing openings and middlegames into scripted execution, from chess engines to modern warfare and AI strategy. This endgame-first reasoning treats real-world uncertainty as a series of bottlenecks to skip, rewarding fluency in simulation over engagement with unresolved complexity. The failure mode emerges when actors mistake the ability to model outcomes for proximity to them, abandoning the middlegame where risk, ambiguity, and real consequences still concentrate.
Studio mode focuses on slow, deliberate storytelling with deep craft and novel perspectives, producing rare, timeless content that gains momentum through careful refinement. Newsroom mode prioritizes fast, responsive storytelling, engaging with ongoing conversations and surfacing insights while the moment is alive. The most effective storytellers cycle between the two, letting newsroom responsiveness inform studio work and studio depth amplify newsroom relevance, creating a feedback loop that balances craft with timeliness.
All large companies work remotely - the only people truly working 'in person' at any of these companies are the ones sitting within earshot. Everyone else is collaborating through screens. Hire people with agency who want to work hard. Employees who prefer remote work just want to do their best work in an environment that suits them.
Why do we consistently reward confidence over competence, and how does Dunning-Kruger play out in startups? The best operators don't start confident - they iterate until they actually know what they're doing, and the real confidence follows naturally.
2026.02.09·
investing1012dot0.substack.com·5 min read·TLDR Founders
The old adage was "don't boil the ocean." With artificial superintelligence on the horizon, it's time to retire that phrase entirely. The fear of the future is directly proportional to the smallness of your ambitions. If your plan is to maintain the status quo, AI is terrifying. If your plan is to build something dramatically bigger, it's the best news you've ever gotten. This is Jevons Paradox for everything: when a resource becomes dramatically more efficient, you don't use less of it, you use vastly more.
Foundation model vendors are rapidly verticalizing, compressing the window for Vertical AI companies that rely on short or medium “last-mile” workflows.
Flickr disclosed a data breach stemming from a compromised third-party email service provider, with attackers potentially accessing names, email addresses, usernames, IP addresses, general locations, and platform activity. The company shut down access to the affected system within hours and notified data protection authorities in both Europe and the US. No passwords or financial data were reportedly exposed, though Flickr has urged users to watch for phishing attempts and review account settings.
Microsoft identified a new ClickFix variant dubbed "CrashFix" that uses a malicious Chrome extension impersonating uBlock Origin Lite to deliberately crash victims' browsers, then tricks users into executing clipboard-copied commands via a fake security warning. The attack chain abuses the legitimate Windows utility finger.exe (renamed to ct.exe) to retrieve obfuscated PowerShell payloads, ultimately deploying a Python-based RAT called ModeloRAT that selectively targets domain-joined enterprise systems for further compromise. ModeloRAT beacons to hardcoded C2 servers over HTTP, establishes persistence via registry Run keys and scheduled tasks, and conducts network reconnaissance using native Windows commands like nltest and net use.
BridgePay Networks Solutions has confirmed that an ongoing outage in its payment gateway and other services is caused by a ransomware attack. The company has stated that it is confident that payment details and credit card information have not been compromised. BridgePay has not responded to questions about which ransomware gang is behind the attack.
Researchers reverse-engineered ring-1.io, a prominent game cheat provider, revealing a sophisticated attack chain that replaces EFI boot binaries, injects into Hyper-V via VMEXIT hooks, and uses EPT-based memory redirection to hide malicious code from kernel-level anti-cheat systems. The implant operates across three privilege boundaries — guest user mode, guest kernel mode, and VMX root — using cloned page tables, CR3 spoofing, and shadow pages filled with 0xCE bytes to evade detection. The analysis also outlines multiple detection strategies, including shadow PML4E scanning, Intel Processor Trace analysis, and RWX anomaly detection under HVCI, emphasizing that Secure Boot enforcement would prevent the entire attack chain.
Researchers discovered that AI IDEs forked from VSCode, including Cursor, Windsurf, and Google Antigravity, inherited hardcoded extension recommendation lists pointing to Microsoft marketplace namespaces that were unclaimed on OpenVSX, the open-source alternative these IDEs actually use. An attacker could have registered these unclaimed namespaces and uploaded malicious extensions that the IDEs would proactively recommend to millions of developers based on file types or installed software. The researchers preemptively claimed the vulnerable namespaces and coordinated with the Eclipse Foundation and vendors to remediate, noting that over 1,000 developers installed their inert placeholder extensions simply because their IDE recommended them.
Static credentials are prone to leakage, can cause significant damage, and can be difficult to rotate. To prevent this, many organizations adopted interim secrets management solutions, such as AWS Secrets Manager and HashiCorp Vault. In the long term, organizations should shift to modern, short-lived credentials, such as managed identities (e.g., AWS roles), federated identities, and application-specific methods for Kubernetes and AI agents.
When G2 analyzed 14 enterprise password managers, the results weren't even close [10]. Bitwarden scored 98/100 on user satisfaction, way above Keeper (92), 1Password (66), and LastPass (58). Bitwarden also had 29% faster ROI, with 83% of Bitwarden customers live in <30 days.
Hermes is a Linux-only Mythic C2 agent written in Python, featuring 18 built-in commands for reconnaissance, file operations, and shell execution. The agent communicates via HTTP, using Mythic's EKE + AES encryption, and can be deployed as a Python script or a PyInstaller binary. It supports core post-exploitation tasks, including process listing, network enumeration, file transfer, and directory manipulation on Linux targets.
2026.02.09·
github.com·1 min read·GitHub repository·TLDR InfoSec
SharePointDumper is a PowerShell-based extraction and auditing utility that enumerates SharePoint sites a user has access to via Microsoft Graph and downloads files via SharePoint.
The Electronic Frontier Foundation launched its "Encrypt It Already" campaign, urging major tech companies such as Bluesky, Google, and Ring to fulfill their promises to implement end-to-end encryption by default across their platforms. The initiative highlights that many companies either offer E2EE as an opt-in feature or have delayed rolling it out entirely, leaving users exposed, particularly as AI agents increasingly access sensitive communications with less human oversight. EFF noted that several targeted companies are "highly likely" to enable these features within the year, but stressed that default-on encryption remains critical since most users never change default settings.
The author received a LinkedIn post about a freelance opportunity to evaluate the codebase of a real estate tech platform with a legitimate-looking platform and a generous budget. The author accepted the offer and was given access to the codebase, but missed red flags such as a missed call with their Tech Manager, suspicious git history, and suspicious indicators in their contact's LinkedIn profile. Upon reviewing the repository, the author found an npm post-install script that would download a C2 backdoor, exfiltrate sensitive files, and capture keystrokes.
Anthropic researcher Nicholas Carlini used 16 parallel Claude Opus 4.6 agents coordinating via a shared Git repository to produce a 100,000-line Rust-based C compiler capable of compiling a bootable Linux kernel across x86, ARM, and RISC-V architectures. The two-week, $20,000 experiment achieved a 99% pass rate on the GCC torture test suite but hit a practical ceiling around 100,000 lines where new fixes frequently broke existing functionality, suggesting current limits for autonomous agentic coding. The researcher noted significant human scaffolding was required — including custom test harnesses, context-aware output filtering, and time-boxing — raising concerns about developers deploying AI-generated software they have never personally verified.
Apple is reportedly developing support for AI chatbot apps like ChatGPT, Gemini, and Claude within CarPlay, potentially allowing drivers to interact with third-party AI assistants alongside Siri.
Malicious npm and PyPI packages published through compromised official dYdX accounts exfiltrated wallet seed phrases and deployed a RAT backdoor, marking the third supply chain attack targeting the decentralized exchange.
A.Team builds agentic INTELLIGENCE that compresses two weeks of analysis into 9 minutes. We unify fragmented data into a single SELF-IMPROVING SYSTEM that embeds in your existing tools and learns from every campaign decision. 48 hours to first insight. 90 days to measured value.
YouTube is expanding Super Bowl advertising beyond traditional 30-second spots by surrounding the event with creator-driven content and live activations. The platform leverages creator partnerships, pre-game events like a flag football game, and sponsorship packages to embed brands into cultural moments. AI-powered campaigns help advertisers reach new audiences, while AdBlitz amplifies exposure, with 45% of viewers seeing content only on YouTube.
Amazon's ad business is growing despite heavy AI spending, raising questions about future returns. Q4 revenue rose 14% to $213.4B, ad revenue grew 22% to $21.3B, and AWS sales increased 24% to $35.6B. Prime Video now has 315M ad-supported viewers, up from 200M, and Thursday Night Football averaged 15M viewers per game, up 16%. Capital spending in 2026 is planned at $200B, about 50% higher than last year, while profit per dollar remains lower than rivals.
The intent economy is a system where marketing responds directly to real-time consumer intent instead of relying on assumptions or campaigns. AI detects needs, acts immediately, and learns from every interaction. The four-part model starts with relevance, which ensures a product is considered. Distinctiveness makes it noticed. Differentiation provides a reason to choose it. Default makes it the automatic choice and allows continuous optimization.
First, set up a subscriber lifecycle automation to check in at key intervals, like 90, 180, 270, and 365 days. Add personalization over time. Second, implement a sunset policy that automatically unsubscribes contacts who show no engagement. Together, these steps protect deliverability, recover potential revenue from partially engaged users, and remove the burden of manual list management.