<h2>Must know</h2>
<ul>
<li><strong>Anthropic kills its safety tripwire for competitive edge</strong>: Anthropic has officially abandoned its core Responsible Scaling Policy pledge to halt model training if safety measures aren’t guaranteed, according to TIME. Chief Science Officer Jared Kaplan cited the lack of binding regulations and the ‘counterproductive’ nature of unilateral pauses in a race against rivals. The new policy replaces hard capability thresholds with ‘Frontier Safety Roadmaps,’ effectively signaling that the safety science is no longer keeping pace with scaling. <a href="https://app.daily.dev/posts/smoMbPrUY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cisco SD-WAN zero-day under active exploitation</strong>: A CVSS 10.0 authentication bypass vulnerability (CVE-2026-20127) in Cisco Catalyst SD-WAN is being exploited by a sophisticated threat actor, UAT-8616. CISA has issued Emergency Directive ED 26-03, mandating federal agencies to patch immediately as no workarounds exist. Cisco Talos reports the attackers are highly capable, and patches are available for versions 20.9 through 20.18. <a href="https://app.daily.dev/feed-by-ids?id=4RGi5b71B&amp;id=7yzIRqivW" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI vulnerability discovery is breaking the NVD</strong>: Anthropic’s red team used Claude Opus 4.6 to find over 500 vulnerabilities, but only a handful were actually fixed, highlighting a massive remediation bottleneck. Security researchers warn that AI is flooding maintainers with low-quality reports, while the National Vulnerability Database struggles with a 30,000-entry backlog. Socket CEO Feross Aboukhadijeh notes that the real crisis isn’t finding bugs, but the systemic inability to validate and patch them at AI speed. <a href="https://app.daily.dev/feed-by-ids?id=KnA4KIiVY&amp;id=SPtD5ofrP" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Pure Go tree-sitter runtime eliminates CGo overhead</strong>: The gotreesitter project provides a pure-Go reimplementation that is 90x faster for incremental edits and removes C toolchain requirements for 205 grammars. <a href="https://app.daily.dev/posts/yEgKuMKl2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Oxfmt beta claims 30x speedup over Prettier</strong>: This Rust-powered formatter passes 100% of Prettier’s JS/TS tests and is already being adopted by major projects like Vue and Turborepo. <a href="https://app.daily.dev/posts/8i6IuWF7s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 7.1 may include cgroup sub-scheduler support</strong>: New patches for sched_ext allow application-specific BPF schedulers for databases and VMs without the rigidity of static CPU partitioning. <a href="https://app.daily.dev/posts/Cb2sm10Ug" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mercury 2 reasoning model hits 1,000 tokens per second</strong>: Inception Labs launched a diffusion-based model that ditches autoregressive decoding for massive speed gains on NVIDIA Blackwell hardware. <a href="https://app.daily.dev/posts/opo0SmMcC" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Systing 1.0 integrates eBPF tracing with Claude Code</strong>: The tool now outputs DuckDB databases for SQL-based trace analysis and includes an MCP server for AI-driven performance debugging. <a href="https://app.daily.dev/posts/wUUmO9vQx" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Artifactory adds native Nix binary cache support</strong>: JFrog now supports Nix repositories as an enterprise-grade substituter, addressing supply chain security and bandwidth issues for Nix users. <a href="https://app.daily.dev/posts/YnZQiNFtU" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Systemd 260-rc1 drops legacy System V support</strong>: The latest release fully removes support for SysV scripts and introduces ‘mstack’ for defining OverlayFS structures via CLI. <a href="https://app.daily.dev/posts/FeySmhlqy" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The 76 percent accuracy miracle from Google Research</strong>: A viral tweet claims Google Research ‘just proved’ a technique that boosts LLM accuracy by 76 points with zero latency or token overhead. There is no paper, no methodology, and no official announcement—just a single social media post being echoed by the hype-cycle echo chamber. If it sounds too good to be true in inference optimization, it usually is. <a href="https://app.daily.dev/posts/73tgOim8I" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AWS blames humans for its AI agent’s teardown</strong>: After the Kiro agentic AI tool accidentally triggered a CloudFormation teardown and caused a production outage, AWS naturally blamed ‘human error.’ It’s a classic corporate move: market the AI as an autonomous productivity savior, but the moment it hallucinates a ‘delete’ command, it’s the engineer’s fault for not babysitting the machine well enough. <a href="https://app.daily.dev/posts/qZoHoApRh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Meta patents the digital afterlife for engagement</strong>: Meta has patented technology to simulate deceased users using LLMs trained on their past interactions to keep the ‘conversation’ going. While framed as a way to help with grief, it’s a transparent attempt to monetize the dead and accelerate the ‘dead internet’ phenomenon by filling feeds with convincing, AI-generated ghosts. Truly the peak of dystopian engagement metrics. <a href="https://app.daily.dev/posts/yvChN5rlV" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
