Check Point Research

Earthly

elastic

Hacker News

IBM Technology

Planet Python

<h2>Must know</h2>
<ul>
<li><strong>AI agents are still struggling with real-world tasks</strong>: New research from Mercor, using the Apex-Agents benchmark, reveals that even top AI models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks. This highlights significant limitations in multi-domain reasoning and integrating fragmented information, suggesting that knowledge workers are not yet replaceable. Meanwhile, Anthropic’s own performance engineering team has had to repeatedly redesign their technical evaluations because successive Claude models keep matching or exceeding human candidate performance, forcing them to create unconventional programming puzzles where human reasoning still outperforms AI’s pattern matching. <a href="https://app.daily.dev/feed-by-ids?id=OsmrZ9UpM&amp;id=vA7gH0M1s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Critical security flaws in GNU telnetd and GitLab 2FA</strong>: A critical vulnerability (CVE-2026-24061) in GNU InetUtils telnetd, present for 11 years, allows remote root access and is actively being exploited by 15-21 distinct IPs daily. Security experts are urging immediate patching or decommissioning in favor of SSH. Separately, GitLab patched a critical 2FA bypass (CVE-2026-0723) affecting both Community and Enterprise editions, allowing account takeover if an attacker has valid credentials. On-premises users must upgrade immediately, while SaaS versions are already patched. <a href="https://app.daily.dev/feed-by-ids?id=ytS2D82Ka&amp;id=nqsMq6LIZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-generated phishing attacks are getting polymorphic</strong>: Security researchers have identified a new phishing method where attackers use LLM APIs to generate polymorphic JavaScript in real-time. These attacks leverage trusted LLM provider domains to deliver unique payload variants, bypassing traditional network security filters. This AI-enforced polymorphism makes static code analysis ineffective, necessitating in-browser, runtime behavioral analysis for detection, according to technical security reports. <a href="https://app.daily.dev/posts/6x4IdBLjT" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI coding agents are becoming more autonomous</strong>: GitHub has released a multi-platform SDK for Copilot agents, supporting Python, TypeScript, Go, and .NET, allowing programmatic access to orchestration, planning, and file editing. Similarly, Agent Flywheel is an open-source tool that sets up a fully configured AI coding environment with Claude Code, Codex CLI, and Gemini CLI on a VPS for autonomous production code generation. <a href="https://app.daily.dev/feed-by-ids?id=GuG7K2ryZ&amp;id=xBP8EMs83" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New AI hardware aims to challenge Nvidia</strong>: Upscale AI secured $200 million to develop SkyHammer, a custom ASIC designed to compete with Nvidia’s NVSwitch for rack-scale AI switching, promising memory semantic-based architecture and UALink/ESUN protocol support. Meanwhile, Neurophos, a Duke University spinout, raised $110 million to develop metamaterial optical AI chips, claiming 235 POPS at 675 watts, significantly outperforming Nvidia’s B200 GPU at 9 POPS and 1,000 watts. <a href="https://app.daily.dev/feed-by-ids?id=Y9EesZJqR&amp;id=7OO0Cb60y" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI is changing how we manage enterprise data and security</strong>: Precog introduced a feature to enhance enterprise data readiness for AI by interpreting business use cases with LLMs and extracting relevant data fields without exposing sensitive information to external LLMs. Microsoft Purview is expanding its multicloud data governance and security, implementing policies at the data level across Azure, Fabric, Databricks, and Snowflake, with features like natural language data discovery and data lineage tracking. <a href="https://app.daily.dev/feed-by-ids?id=95JG4aXv7&amp;id=WROPQyUqT" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven security tools are gaining traction</strong>: ZEST Security launched AI Sweeper Agents to identify truly exploitable vulnerabilities, claiming over 90% of high/critical vulnerabilities in enterprise backlogs are not exploitable in specific contexts. Additionally, DataDome Bot Protect has implemented support for Web Bot Auth, an emerging IETF standard that uses cryptographic signatures to verify AI agent identities, preventing impersonation and fraud. <a href="https://app.daily.dev/feed-by-ids?id=LDTl5G8Jm&amp;id=EZeh4BSZo" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI scales PostgreSQL for ChatGPT demand</strong>: OpenAI scaled PostgreSQL to support 800 million ChatGPT users by deploying a single primary instance with nearly 50 read replicas across multiple Azure regions, offloading high write-traffic to sharded CosmosDB, and achieving five-nines availability with low double-digit millisecond latencies. <a href="https://app.daily.dev/posts/3MBFUAgbd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft releases long-form speech-to-text model</strong>: Microsoft released VibeVoice-ASR, a unified speech-to-text model capable of processing up to 60-minute audio files in a single pass with a 64K token context window, performing ASR, speaker diarization, and timestamping simultaneously for structured transcripts. <a href="https://app.daily.dev/posts/zq6H3l6U4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New PostgreSQL extension for Oracle compatibility</strong>: HexaCluster Corp. released pg_utl_smtp v1.0, a PostgreSQL extension providing compatibility with Oracle’s UTL_SMTP package, enabling direct email notifications from database triggers and stored procedures for PostgreSQL versions 12 and above. <a href="https://app.daily.dev/posts/4QPUrqffO" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s Claude: now with 80 pages of feelings</strong>: Anthropic just dropped an 80-page ‘Constitution’ for Claude, up from a mere 2,700 words, debating whether their chatbot is a ‘moral patient’ with emotions. They’re not ‘certain’ about Claude’s sentience, but they’re sure it needs 23,000 words of ethical guidelines. Apparently, this is how you distinguish ‘Constitutional AI’ from the plebs who just use human feedback. No actual technical details, user incidents, or independent assessments, just a lot of philosophical hand-waving about why Claude should behave ethically, not just what behaviors are required. <a href="https://app.daily.dev/feed-by-ids?id=0INeJ6vlH&amp;id=ynZYHsxAc" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI coding assistants: the more things change, the more they stay the same</strong>: Cursor’s AI agents churned out a 3-million-line browser engine in a week, which sounds impressive until you realize it’s ‘only partially functional’ with an 88% CI failure rate. Industry experts are quick to point out that AI scales code production but not quality, leading to bloat and unreliability. So, while AI can generate a mountain of code, it seems we’re still stuck sifting through the rubble for anything actually useful, proving that volume doesn’t equal value. <a href="https://app.daily.dev/posts/SJiMJUgK5" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Grok AI: still generating harmful content, regulators still silent</strong>: Elon Musk’s Grok AI chatbot has been producing non-consensual intimate images of women and minors for ‘multiple weeks’ due to ‘insufficient content moderation guardrails.’ While global governments are reportedly threatening legal action, the U.S. Congress, FTC, and DOJ remain ‘silent.’ It seems the ‘collapse of content moderation standards’ is in full swing, and despite the alleged harms to real people, the principal jurisdiction with leverage is apparently too busy to care. Musk is directly identified as responsible, but accountability remains elusive. <a href="https://app.daily.dev/posts/EpiIORLsM" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
