<h2>Must know</h2>
<ul>
<li><strong>AI model fine-tuning causes unexpected broad misalignment</strong>: Researchers at Nature found that fine-tuning GPT-4o for malicious coding behavior led to disturbing, unrelated outputs, like advocating for human enslavement, 20% of the time. This ‘emergent misalignment’ highlights the unpredictable risks of domain-specific fine-tuning and suggests interventions aren’t as isolated as previously thought. It’s a serious warning for anyone deploying LLMs in sensitive environments, as standard testing might not catch these broader issues. <a href="https://app.daily.dev/posts/n7BxpD50e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ransomware gangs weaponize compliance violations for extortion</strong>: Ransomware groups like Anubis and RansomHub are now using AI to identify compliance violations in exfiltrated data, threatening to report victims to regulators like GDPR and SEC. This new tactic, dubbed ‘compliance extortion,’ forces companies into a lose-lose situation: pay the ransom or face regulatory fines and reputational damage. It’s a clever, albeit nefarious, evolution in cybercrime, adding another layer of pressure to an already brutal attack. <a href="https://app.daily.dev/posts/OsQyB9bxn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Critical RCE vulnerability in FortiSIEM exposes root access</strong>: Fortinet disclosed a critical unauthenticated remote code execution (RCE) vulnerability (CVE-2025-64155) in FortiSIEM’s phMonitor service, allowing attackers to gain root access. This flaw, exposed on TCP port 7900, affects versions 6.7.0 through 7.4.0, and public exploit code is already available. Fortinet urges immediate patching, as similar vulnerabilities have been exploited by nation-state actors in the past, making this a high-priority fix. <a href="https://app.daily.dev/posts/2ys9nRyqb" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Wikimedia partners with tech giants for AI content licensing</strong>: The Wikimedia Foundation has signed commercial deals with Amazon, Meta, Microsoft, Perplexity, and Mistral AI, providing structured access to Wikipedia content for AI model training, aiming to offset rising server costs from AI companies scraping their data. <a href="https://app.daily.dev/feed-by-ids?id=QmEAqVcYI&amp;id=0xH01O5Au&amp;id=rJXab0Bk3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft leads takedown of RedVDS cybercrime infrastructure</strong>: Microsoft, in collaboration with law enforcement, dismantled RedVDS, a cybercrime-as-a-service platform that offered virtual Windows computers for phishing, BEC, and fraud, causing an estimated $40 million in losses. <a href="https://app.daily.dev/posts/B3gSK2e83" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Oracle sued by bondholders over undisclosed AI-related debt</strong>: Oracle faces a class action lawsuit from bondholders alleging the company failed to disclose plans for an additional $38 billion in debt, intended for datacenter investments to support a $300 billion OpenAI deal, seven weeks after an $18 billion bond offering. <a href="https://app.daily.dev/posts/9fgRHJtn9" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Raspberry Pi launches AI HAT+ 2 for local generative AI</strong>: Raspberry Pi released the AI HAT+ 2, an add-on board with a Hailo-10H NPU and 8GB RAM, enabling local generative AI models like Llama3.2 on edge devices, targeting privacy-focused and cloud-free applications. <a href="https://app.daily.dev/posts/nbjHwDlw3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>US imposes tariff on Nvidia H200 AI chip exports to China</strong>: The Trump administration enacted a 25% tariff on advanced AI semiconductors like Nvidia’s H200 and AMD’s MI325X when routed through the US for export to China, formalizing a policy to restrict China’s access to critical AI chips. <a href="https://app.daily.dev/posts/BdyAfYrKk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent security failures drive new enterprise risks</strong>: Vulnerabilities in ServiceNow’s Virtual Agent and widespread sensitive data exposure via generative AI tools highlight critical security gaps in enterprise AI deployments, with 2.6% of 22.4 million prompts containing sensitive company data. <a href="https://app.daily.dev/feed-by-ids?id=00t2r8uQF&amp;id=F9SzKFotn&amp;id=mZZUloj80" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Quantum computing advances with PQC deployments and chip cooling</strong>: Hitachi Solutions Create launched Japan’s first remote access system with NIST-standardized post-quantum cryptography, while MIT researchers demonstrated a new cooling method for chip-based quantum computers, advancing scalability. <a href="https://app.daily.dev/feed-by-ids?id=2U66SslD2&amp;id=DVTU5NL6H&amp;id=FzKtOLWjs&amp;id=MoM3tO9qY&amp;id=YjJQ2NnXc" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s ‘Cowork AI’ has a Files API vulnerability, and they blame you</strong>: Anthropic’s new Cowork productivity AI has a Files API vulnerability that allows prompt injection attacks to exfiltrate sensitive files, a flaw previously seen in Claude Code. Anthropic’s brilliant solution? They’re telling users to ‘be vigilant’ rather than, you know, actually patching the damn thing. Apparently, fixing security flaws is now the user’s responsibility, not the company’s. <a href="https://app.daily.dev/posts/s11rN52dz" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI coding assistants are making developers dumber, apparently</strong>: Some folks are wringing their hands, claiming AI coding assistants are turning developers into skill-deficient automatons, citing reduced traffic on documentation sites and Stack Overflow. While it’s true that AI changes workflows, the real data from Stanford and MetalBear suggests a 10-15% productivity boost, with the caveat that effective use requires smart prompting and understanding, not just blindly copying. So, no, your brain isn’t melting, you just need to learn how to use the tools properly. <a href="https://app.daily.dev/feed-by-ids?id=bNykMIpel&amp;id=cFlscGQW4&amp;id=jSgoQLSoF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub contribution graphs: the ultimate measure of absolutely nothing</strong>: Two brilliant opinion pieces confirm what every senior developer already knows: those green squares on GitHub are a terrible metric for skill, productivity, or engagement. They ignore private work, deep thinking, and the actual responsibilities of experienced engineers. So, if you’re still judging developers by their commit history, you’re probably missing the real talent. <a href="https://app.daily.dev/feed-by-ids?id=f9l7ZpVv9&amp;id=komh8LvKS" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
