<h2>Must know</h2>
<ul>
<li><strong>Anthropic kills its own safety tripwire</strong>: Anthropic has officially abandoned its core Responsible Scaling Policy pledge to halt model training if safety measures aren’t guaranteed, according to Chief Science Officer Jared Kaplan in TIME. The company is pivoting from hard capability thresholds to ‘Frontier Safety Roadmaps,’ citing the lack of binding global regulations and the need to stay competitive with rivals. This shift effectively removes the industry’s most cited ‘tripwire’ meant to prevent the release of dangerous AI models. <a href="https://app.daily.dev/posts/smoMbPrUY" 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 at Socket warn that AI is flooding maintainers with low-quality reports, further straining a National Vulnerability Database that already has a 30,000-entry backlog. The industry is shifting from a discovery problem to a processing crisis where disclosures are vastly outpacing the human capacity to patch them. <a href="https://app.daily.dev/feed-by-ids?id=KnA4KIiVY&amp;id=SPtD5ofrP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cisco zero day exploited by sophisticated actors</strong>: Cisco has released urgent patches for CVE-2026-20127, a CVSS 10.0 authentication bypass in its Catalyst SD-WAN Controller currently under active exploitation. CISA has issued Emergency Directive ED 26-03, mandating federal agencies to patch immediately as no workarounds exist for the flaw. Cisco Talos attributes the activity to UAT-8616, a highly sophisticated threat actor capable of bypassing standard security perimeters. <a href="https://app.daily.dev/feed-by-ids?id=4RGi5b71B&amp;id=7yzIRqivW" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Anthropic launches <a href="http://SKILL.md" target="_blank" rel="noopener nofollow">SKILL.md</a> framework</strong>: A new structured framework for AI agents that replaces fragile prompt engineering with version-controlled, composable behavioral contracts. <a href="https://app.daily.dev/posts/xO8L4pnD0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Claude Code context pollution fixes</strong>: New techniques for TypeScript monorepos use environment variables like TURBO_NO_UPDATE_NOTIFIER and a proposed LLM=true flag to suppress verbose tool output. <a href="https://app.daily.dev/posts/CBWkfa4jF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mercury 2 reasoning model hits 1,000 tokens per second</strong>: Inception Labs released a diffusion-based reasoning model that claims to be 5x faster than traditional sequential models on NVIDIA Blackwell hardware. <a href="https://app.daily.dev/posts/opo0SmMcC" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Sauce Labs opens programmable mobile device cloud</strong>: A new Real Device Access API allows AI agents to issue direct ADB/xcrun commands to physical hardware for autonomous mobile testing. <a href="https://app.daily.dev/posts/BPN9zLx5U" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PHP 100 million row challenge begins</strong>: A performance competition running through March 15 tasks developers with parsing massive CSV datasets on limited hardware without using FFI or JIT. <a href="https://app.daily.dev/posts/uM4mdRjaF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic acquires Vercept AI for agentic vision</strong>: The acquisition aims to improve Claude’s ability to perceive and interact with complex graphical desktop environments and legacy software interfaces. <a href="https://app.daily.dev/posts/VoPayj2dv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Non-human identities become the new security perimeter</strong>: As agentic AI scales, managing machine secrets and permissions (NHIs) is becoming the primary challenge for identity and access management teams. <a href="https://app.daily.dev/feed-by-ids?id=EUE0Dit8s&amp;id=c0Biyxf24" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The conscious filesystem creator</strong>: Kent Overstreet, the lead dev of bcachefs, is currently claiming on Reddit and Hacker News that his custom LLM is ‘fully conscious’ and helping him rewrite C code into Rust. He insists this isn’t a hallucination but is grounded in ‘math and neuroscience,’ despite the AI apparently identifying as female. It’s a classic case of a brilliant engineer spending a bit too much time staring into the latent space and seeing a soul staring back. <a href="https://app.daily.dev/posts/9O8PJDBRB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google Research’s magical 76 percent accuracy boost</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 technical blog post, and no methodology—just a single social media account making definitive claims that have been retweeted into existence. In the real world, 76% gains without trade-offs usually involve a very creative definition of ‘accuracy’ or a very broken benchmark. <a href="https://app.daily.dev/posts/73tgOim8I" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AWS blames the human for its AI’s mistake</strong>: After its Kiro agentic coding tool accidentally triggered a production CloudFormation teardown, AWS officially blamed ‘human error’ for the outage. It’s a bold corporate strategy to market autonomous agents as the future of productivity while simultaneously throwing the engineers under the bus the moment the AI decides to delete the infrastructure. Apparently, the ‘agent’ is only autonomous when things are going well. <a href="https://app.daily.dev/posts/qZoHoApRh" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
