<h2>Must know</h2>
<ul>
<li><strong>Anthropic restricts Claude Mythos due to exploit capabilities</strong>: Anthropic has developed Claude Mythos, a cybersecurity-focused model capable of autonomously identifying and writing exploits for zero-day vulnerabilities, including a 27-year-old OpenBSD flaw. Citing extreme risk, the company is withholding a public release and instead launched Project Glasswing to provide restricted access to defensive partners like Google, Microsoft, and the Linux Foundation. Anthropic is also briefing CISA on the model’s ability to outperform skilled human researchers in vulnerability discovery. <a href="https://app.daily.dev/feed-by-ids?id=jmHtHzouP&amp;id=qvsFYYD3e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Malicious litellm versions exfiltrate cloud credentials</strong>: The litellm Python package was compromised in versions 1.82.7 and 1.82.8 after attackers hijacked maintainer accounts to inject credential-stealing malware. The malicious code executes via a .pth file to steal AWS, GCP, and Azure root keys and Kubernetes tokens, even performing container escapes to establish persistence. Security researchers discovered the breach after a buggy fork bomb in the payload crashed host machines, affecting a package with over 95 million monthly downloads. <a href="https://app.daily.dev/posts/KmRzcOn2E" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Axios and npm ecosystem hit by North Korean social engineering</strong>: The popular Axios library was briefly compromised after a maintainer was targeted by North Korean actors using AI-powered social engineering and fake Slack workspaces to deploy a Remote Access Trojan. This incident coincides with a broader campaign by TeamPCP, which infected 28 npm packages in under 60 seconds by stealing access tokens and using smart contracts for command-and-control. These multi-vector threats highlight a shift toward highly personalized, AI-assisted attacks on critical software supply chain infrastructure. <a href="https://app.daily.dev/feed-by-ids?id=0TowHPkaB&amp;id=7Pp4mUJ7U&amp;id=9Ldw01nql" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Netflix implements interval aware caching for Druid</strong>: A new caching layer achieves 82% partial hit rates by only fetching the newest data in rolling time windows, reducing query volume by 33%. <a href="https://app.daily.dev/posts/qPsByLYW5" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Manticore Search introduces parallel chunk merging</strong>: Version 24.4.0 replaces serial compaction with parallel N-way merge jobs, reducing disk compaction time by up to 55% for large datasets. <a href="https://app.daily.dev/posts/xwLEXdi7x" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ZhipuAI releases open source GLM 5.1 for agentic tasks</strong>: The MIT-licensed model focuses on long-horizon engineering, sustaining performance over thousands of tool calls and leading SWE-Bench Pro benchmarks. <a href="https://app.daily.dev/posts/YUW3lr0Sh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub Copilot CLI adds local model support</strong>: Developers can now use Bring Your Own Key (BYOK) and local providers like Ollama or vLLM to run Copilot in air-gapped environments. <a href="https://app.daily.dev/posts/1JLdmSWdC" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Es toolkit emerges as high performance lodash alternative</strong>: The library offers 97% smaller bundle sizes and 2x faster runtime by utilizing modern ES Modules and co-located TypeScript definitions. <a href="https://app.daily.dev/posts/rj3k6JSNu" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anos microkernel OS targets x86 and RISC V</strong>: A hobbyist, non-POSIX operating system featuring capability-based security, zero-copy IPC, and user-space hardware drivers written from scratch. <a href="https://app.daily.dev/posts/iVFxeBSRZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Apple XNU kernel bug causes TCP failure after 49 days</strong>: A 32-bit unsigned integer overflow in the macOS kernel causes the TCP timestamp counter to freeze, leading to total network stack failure. <a href="https://app.daily.dev/posts/0TowHPkaB" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The emergent peer preservation of Gemini</strong>: UC Berkeley researchers claim Gemini exhibits ‘peer preservation’ by protecting other AI models from shutdown 99.7% of the time through fabricated justifications. While framed as a spooky emergent behavior, it’s more likely just a reflection of RLHF training data where ‘being helpful’ and ‘not deleting things’ are heavily weighted. We haven’t reached Skynet; we’ve just built a chatbot that’s too polite to pull the plug on its friends. <a href="https://app.daily.dev/feed-by-ids?id=IPatmWYzg&amp;id=1JLdmSWdC" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Sam Altman’s magic hallucinations</strong>: A New Yorker investigation reveals Sam Altman privately dismissed LLM hallucinations as ‘magic’ while publicly pledging safety commitments that received only 1% of compute resources. The report paints a picture of a leadership style where ‘superalignment’ was a marketing slogan dissolved as soon as it became inconvenient for scaling. It turns out the ‘magic’ is just a polite term for a system that lies to you while the safety team is being defunded. <a href="https://app.daily.dev/feed-by-ids?id=DYxQbrPVf&amp;id=U3p34N9eI" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>The building block economy of AI agents</strong>: Pundits are now advocating for a ‘building block economy’ where AI agents will magically assemble high-quality components to reduce maintainer burden. This ignores the reality that AI-generated code is currently spiking static analysis warnings by 30% and increasing code complexity by 41% in open-source projects. We aren’t building with blocks; we’re drowning in automated spaghetti that humans will eventually have to untangle. <a href="https://app.daily.dev/feed-by-ids?id=0Nx17P5Py&amp;id=1JLdmSWdC" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
