Business, Deals & Funding
Ars Technica AI
Private security firms will soon be allowed to hack overseas cybercriminals
The Trump administration has issued a National Security Presidential Memorandum authorizing private security firms to conduct offensive cyber operations, including cyberattacks and surveillance, against overseas criminal organizations that target US persons, organizations, or government entities. This marks the first time the federal government has authorized private companies to perform such operations. Eligible targets include ransomware groups, sextortion schemes, phishing campaigns, and financial fraud operations run by foreign transnational criminal organizations. Participating firms must be vetted by the Departments of Justice and Homeland Security, and operations cannot result in loss of life, serious injury, or rise to the level of armed attack under international law. Security researcher Kevin Beaumont noted that while hacking ransomware groups has merit and already occurs info…
Why it matters
This policy represents a significant and potentially dangerous expansion of offensive cyber capabilities into the private sector. While the frustration with overseas cybercriminals is understandable, privatizing cyberattacks raises serious concerns about accountability, escalation, and unintended consequences. The memo's vague language—not explicitly ruling out DDoS attacks or ransomware-like encryption tactics—is troubling. Kevin Beaumont's point about perverse incentives is particularly sharp…
Claude Code Changelog
v2.1.231
Fixed a bug where MCP OAuth sign-in would fail with a redirect URI mismatch error when connecting to servers that use a pre-registered OAuth client, such as Slack.
Why it matters
This is a minor but important bug fix for users trying to authenticate with MCP servers via OAuth. Redirect URI mismatches are a common and frustrating OAuth integration issue, so resolving this for pre-registered OAuth clients like Slack improves the reliability of MCP server connections.
Cursor Blog
Cursor earns AIUC-1 certification for agent security and reliability
Cursor announced it has earned AIUC-1 certification, a new standard for AI agent security, safety, and reliability. The certification combines an independent audit of organizational controls (conducted by Schellman) with adversarial testing of Cursor's agents across thousands of scenarios. AIUC-1 was developed with input from over 100 Fortune 500 CISOs and organizations like MITRE, the Cloud Security Alliance, and Stanford researchers, translating frameworks like NIST AI RMF and MITRE ATLAS into testable requirements for live AI systems. The testing covered areas including secrets protection, secure code generation, MCP security, and agent identity/permissions. Cursor's safeguards—including rules, hooks, and Auto-review—held across both benign and adversarial conditions. The certification requires quarterly re-testing and annual full audits, with the standard itself updated quarterly to…
Why it matters
This is a meaningful step toward establishing accountability standards for AI coding agents, addressing a real gap where traditional security certifications cover data governance but not agent behavior under adversarial conditions. The involvement of credible organizations (MITRE, CSA, Stanford) and an accredited auditor lends legitimacy. The quarterly re-evaluation requirement is particularly smart, acknowledging that AI agent capabilities and risks evolve rapidly. However, this is also clearl…
DeepMind Blog
Introducing Gemini 3.7 Flash
Google DeepMind announced Gemini 3.7 Flash, their latest workhorse model optimized for coding and agents, released just three weeks after Gemini 3.6 Flash. The model delivers substantial improvements in software engineering, knowledge work, and web development, with notable gains in benchmarks like FrontierCode 1.1 Main (43.6% vs 34.4%) and DeepSWE v1.1 (65.3% vs 49.0%). It generates more functional layouts and feature-complete apps in fewer prompts. The model is offered at an introductory price of half the original 3.6 Flash cost per million tokens. Google credits developer feedback and algorithmic innovations for the rapid improvements.
Why it matters
The rapid three-week iteration cycle from 3.6 to 3.7 Flash is impressive and signals an aggressive pace of model development at Google. The benchmark improvements are significant — particularly the jump from 49.0% to 65.3% on DeepSWE — suggesting meaningful real-world coding capability gains rather than marginal improvements. The halved pricing is a strong competitive move, likely aimed at maintaining developer mindshare against competitors like Claude and GPT models. However, the article is cl…
Guardian AI
Unemployed young people to join AI boot camps to get job-ready
The UK government is launching a pilot scheme offering three-week 'AI boot camps' for unemployed young people and those at risk of unemployment. The program aims to help participants harness AI technology to improve their employability and gain a foothold in the workplace, as part of efforts to address the growing crisis of NEETs (young people not in education, employment, or training). The initiative is notable for using a technology that many consider a potential threat to employment as a tool to combat joblessness.
Why it matters
This initiative reflects a pragmatic but somewhat superficial approach to a deep structural problem. Three weeks of AI training is unlikely to meaningfully transform employment prospects for young people facing systemic barriers to work. While digital literacy and AI familiarity are increasingly valuable, the NEET crisis is driven by complex factors including mental health challenges, inadequate education systems, housing costs, and a labor market that has been hollowing out mid-skill jobs. Tea…
MIT Tech Review AI
Flock is tightening its rules in response to a growing surveillance backlash
Flock, a police-tech company operating 120,000 license plate reader cameras nationwide, is implementing new rules to address growing backlash over surveillance concerns and police abuse. Key changes include requiring officers to enter criminal case numbers before searches, mandatory automatic auditing of suspicious search activity, recommending shorter data retention periods (7 days instead of 30), and allowing departments to restrict other agencies' search purposes. These changes follow a Washington Post investigation finding 46 cases of officers using Flock cameras for unauthorized purposes like stalking, and reports of officers circumventing existing safeguards. At least 30 cities have dropped Flock contracts in the past year, with criticism coming from both civil liberties groups and conservative commentators like Tucker Carlson. However, critics note the new measures contain signif…
Why it matters
This story illustrates a familiar pattern in surveillance technology: a company deploys powerful monitoring tools with minimal safeguards, abuses predictably occur, and then the company introduces reforms that appear meaningful but contain fundamental weaknesses. The fact that officers entered 'hehehe' 20 times as their search justification perfectly encapsulates how toothless self-reported compliance mechanisms are. Flock's reforms are essentially voluntary guardrails — case numbers aren't ver…
NY Times
As Comedians Toy With A.I., Who Will Get the Last Laugh?
Several acts at the Edinburgh Fringe Festival are experimenting with artificial intelligence to test the boundaries of comedy and creativity, exploring how AI tools can be integrated into or challenge traditional comedic performance.
Why it matters
This is a fascinating cultural development worth watching. Comedy has always been a deeply human art form rooted in timing, shared experience, and subversion of expectations, so using AI in this context raises genuinely interesting questions about authorship, authenticity, and what makes something funny. The Edinburgh Fringe is exactly the right venue for this kind of experimentation. However, the real test will be whether AI-assisted comedy can move beyond novelty and gimmickry to produce genu…
OpenAI
The builder’s guide to GPT‑5.6
This OpenAI blog post from August 2026 introduces GPT-5.6 as a new standard for price-performance in AI. It highlights how startups are using smarter model selection and new Responses API capabilities—including programmatic tool calling, multi-agent orchestration, and prompt caching—to build faster, more cost-efficient AI agents. Key claims include that GPT-5.6 Sol at 'low' reasoning effort outperformed GPT-5.5 at 'high' reasoning, and that the smaller models in the 5.6 family (Luna and Terra) can match GPT-5.4/5.5 performance at dramatically lower costs. Startup testimonials from Hex, Hypha, Browser Use, and PlayerZero report massive cost reductions (up to 18x cheaper) while maintaining near-equivalent accuracy. On the BrowseComp benchmark, GPT-5.6 Luna achieved 84.04% accuracy at $1.33 compared to GPT-5.5's 84.36% at $33.27—a roughly 25x cost reduction for essentially the same perform…
Why it matters
This is a well-structured product marketing piece that effectively uses concrete benchmarks and real startup testimonials to make its case. The cost-performance improvements cited are genuinely impressive if accurate—a 25x cost reduction on BrowseComp with near-identical accuracy is remarkable. However, as with all vendor-published guides, the testimonials are cherry-picked success stories, and the benchmarks are chosen to flatter the new models. The article would benefit from discussing failur…
TechCrunch AI
Writer introduces new AI model and upgraded harness to contain token costs
Writer launched Palmyra X6, a new flagship AI model built as a post-training variation on Z.ai's open source GLM-5.2, along with significant upgrades to its agentic harness infrastructure. The company claims the combination can cut customer costs by up to 50% for basic tasks. CEO May Habib emphasized that enterprises are frustrated with rising AI costs and chasing benchmarks, and that harness optimization is a key lever for cost reduction. A Writer research paper found harness efficiency changes reduced costs by an average of 40% across models, often more reliably than model choice alone. The model will be available alongside other models through Writer's model-agnostic platform.
Why it matters
This is a strategically smart move by Writer that addresses a real and growing pain point in enterprise AI: runaway token costs. The emphasis on harness optimization rather than just model performance is a genuinely differentiated approach, and the research backing it up adds credibility. Habib's framing of AI labs as having misaligned incentives around token usage is pointed but fair. However, the 50% cost reduction claim should be scrutinized—'basic tasks' is doing a lot of heavy lifting in t…
The Rundown AI
Grok 4.6 storms the AI frontier
SpaceXAI has launched Grok 4.6, a new AI model that scores 61 on Artificial Analysis' Intelligence Index, placing it just behind Anthropic's Opus 5 (63) and Fable 5 (62), while surpassing OpenAI's GPT-5.6 Sol. The model is priced at $2/$6 per 1M tokens, roughly 60% less than frontier competitors, and outperforms rivals on certain coding and professional/legal benchmarks. Elon Musk has declared it 'objectively No. 1 when considering intelligence, speed & cost' and teased that Grok 4.7 will arrive in 3-4 weeks, claiming it will 'exceed all current models.' The article frames this as a dramatic turnaround for the Grok line, which was previously considered a punchline in the AI industry. The newsletter also includes segments on teaching AI agents to replicate personal writing style, advice from an AI educator on why people misuse AI by asking it to finish tasks rather than prepare for them,…
Why it matters
This article, dated August 2026, presents an interesting trajectory for xAI/SpaceXAI's Grok models, though it should be read with significant caveats. The newsletter relies heavily on Elon Musk's own characterizations of his product as 'objectively No. 1,' which is classic Musk hyperbole — a 61 on the Intelligence Index behind models scoring 62 and 63 is not number one by any objective measure, even when factoring in cost. The framing of Grok going 'from punchline to frontier' is compelling nar…
The Verge AI
Microsoft’s Clippy-like Mico character is no longer the face of Copilot
Microsoft is retiring Mico, the emotive yellow blob avatar that served as the face of Copilot's voice mode since its launch in October 2025. Mico, pitched by former Copilot lead Mustafa Suleyman as a way to give the chatbot an identity, will be moved to Microsoft's Learn Live platform. The change is part of Microsoft's broader effort to merge its Copilot and Microsoft 365 Copilot apps. Mico joins other retired Microsoft virtual helpers like Clippy, Cortana, and Rover. Microsoft says the learnings from Mico about warmth and expressiveness will shape Copilot going forward.
Why it matters
This feels like a predictable outcome for a feature that never quite found its purpose. Mico always seemed like a solution in search of a problem — a cute animated blob that didn't meaningfully improve the chatbot experience. Microsoft has a long history of creating anthropomorphized digital assistants (Clippy, Cortana, Rover) and then quietly shelving them when they fail to resonate. The fact that Mico lasted less than a year suggests it was more of a branding exercise than a genuinely useful…
Claude Code Changelog
v2.1.232
Claude Code v2.1.232 introduces subagent forking as a default feature, where fork-type subagents inherit the full conversation and prompt cache. Non-teammate agent spawns in interactive sessions now run in the background by default. Users can type '@' in the prompt to mention another Claude session by name, and SendMessage now delivers to a bare name matching a live session without requiring confirmation first.
Why it matters
This update focuses on improving multi-agent collaboration and session management in Claude Code. The subagent forking feature being enabled by default suggests it has matured past experimental status, and inheriting the full conversation context and prompt cache should make spawned agents more efficient. The '@' mention syntax and streamlined SendMessage delivery are nice quality-of-life improvements that make inter-session communication more intuitive. These are meaningful workflow improvemen…
Guardian AI
Massachusetts teen accused of killing mother and brother used ChatGPT
A 17-year-old Massachusetts teenager, Arjun Aravind, has been accused of killing his mother and younger brother and is being held without bail. Prosecutors say the case is connected to his use of ChatGPT, with the district attorney stating that Aravind used the internet and AI to search for fantasy stories regarding killing his family. He appeared for arraignment in Concord district court, where a not-guilty plea was entered on his behalf to murder and several additional charges.
Why it matters
This is a deeply tragic case that raises serious questions about the intersection of AI tools and vulnerable individuals, particularly minors. While it would be premature to assign direct causation to ChatGPT for these horrific acts, the reported connection warrants thorough investigation into how AI chatbots may influence or reinforce dangerous ideation in young people. It highlights the urgent need for stronger safeguards, age verification, and content moderation in AI systems, as well as bro…
MIT Tech Review AI
How kids feel about AI, in their own words
MIT Technology Review interviewed kids aged 10 to 18 about their feelings toward AI. The responses were nuanced: many admitted to using AI at least a little, primarily for searching information, schoolwork, and entertainment, consistent with Pew Research data showing 57% of US teens use chatbots for search and 54% for schoolwork. However, reactions ranged from indifference ('bruh,' 'meh') to active resistance due to environmental concerns or fears about creativity and critical thinking. Most kids' first AI encounters came through parents, schools, or tools already embedded in their devices. The authors noted that kids were less worried about AI taking their jobs than about broader societal harm, and many could clearly articulate boundaries around what tasks they would and wouldn't delegate to AI. The article argues that rather than shielding kids from AI, adults should teach them to use…
Why it matters
This is a thoughtful and well-executed piece of journalism that goes beyond the typical moral panic or techno-utopian framing that dominates AI coverage. The decision to let kids speak in their own words is refreshing and reveals a generation that is neither naively enthusiastic nor reflexively fearful—they're pragmatic and discerning. The driving analogy is apt: the answer to a powerful technology isn't avoidance but education. What stands out most is the finding that kids can articulate clear…
OpenAI
Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
OpenAI announced a preview of 'Ultrafast,' a new API service tier that runs GPT-5.6 Sol up to 14 times faster than standard processing, generating up to 750 output tokens per second. Powered by Cerebras hardware, Ultrafast is designed for time-sensitive applications where frontier-level intelligence previously required sacrificing speed. Use cases highlighted include incident response, financial research, real-time customer support and voice AI, commerce personalization, and live research experimentation. Early customers from companies like Jane Street and Podium report that the speed improvement fundamentally changes how they can use AI models in production workflows. The service is launching first in the OpenAI API during a preview period with an initial group of customers, with plans to expand access based on learnings from real production environments.
Why it matters
This is a significant development that addresses one of the most practical barriers to deploying large language models in production: latency. The partnership with Cerebras is particularly interesting, as it suggests OpenAI is willing to look beyond traditional GPU infrastructure to achieve performance breakthroughs. At 750 tokens per second, many applications that previously required smaller, less capable models for speed reasons can now use frontier-level intelligence. The use cases they high…
TechCrunch AI
Databricks wanted to raise $1B, investors wanted $15B. It settled on $5B at a $190B valuation.
Databricks raised $5 billion at a $190 billion valuation after initially planning to raise only $1 billion. CEO Ali Ghodsi said investor interest reached $15 billion after a news report about the fundraise went viral during their conference. The round was led by Coatue, with participation from Blackstone, MGX, T. Rowe Price, and new investor Sixth Street Growth, among about two dozen VCs. Databricks reports $7 billion in annualized run rate revenue growing at 80%, is cash-flow positive, and its cloud data warehouse product alone represents $1.5 billion in run rate growing at 100% year-over-year. Its AI database product Lakebase has hit $100 million revenue run rate since launching in June 2025. Ghodsi justified the large raise by citing expensive AI research costs, multibillion-dollar cloud commitments with major hyperscalers, and an active M&A strategy, including recent acquisitions of…
Why it matters
Databricks' fundraise is a remarkable testament to the current AI investment frenzy, but the underlying fundamentals actually justify much of the enthusiasm. $7 billion ARR growing at 80% while being cash-flow positive is genuinely exceptional for a company at this scale — those are metrics that would command a premium valuation even in a more sober market. The $190 billion valuation implies roughly a 27x revenue multiple, which is high but not absurd for a company growing this fast with positi…
From X/Twitter
- Cactus shipped Needle 2, a 45M-param tool-calling model that fits in a 14MB binary, runs agent sessions in ~28MB of RAM, and trades wins with models 5–70× its size.
- Harrison Chase makes the case that owning your intelligence means three things: an open agent system, a compounding eval loop, and a governed runtime.
- Jason Fried just published the internal 37signals Manager Playbook — covering expectations, leadership philosophy, the "Hell Yeah" standard, and navigating hard conversations.
- Boris Cherny is running a quiet experiment: Claude takes over day-to-day maintenance of production apps via a dedicated Slack channel, with early signs of life.
- Simon Hamp built a fully native Mac desktop app driven by Laravel and Blade — no HTML, CSS, or JS in sight.
- Martin Donadieu dropped his Ahrefs plan and switched to an open-source alternative after the paid tool stopped crawling his site entirely.
- Nader Dabit used the actual X algorithm to build insidetheforyou, a tool that shows how your interactions shape your future feed using real post data.
- Claude Code sessions can now have names and DM each other — spin up `--name backend` and `--name frontend` and let them coordinate.
- Ryan Carson says /goal on Codex is unstoppable — brief praise, no caveats.
- Anthropic is cutting usage limits by 50% on August 19th. Plan accordingly.
- Rakazo launches as an open-source Grok Bot alternative — any LLM, any sandbox, web/desktop/mobile, Apache 2.0.
- Cursor acquires the Firetiger team to build agents that follow code into production and fix what breaks.
- Katie Keith used Claude to migrate off Mailchimp to FluentCRM, ditching a $4,000/year renewal she'd been dreading for years.
- Pieter Levels highlights a DHH tip: hook Claude Code up to your Ubiquiti router to optimize your WiFi.
- Cursor's cloud agents now start 3× faster thanks to pre-built dev environments that spin up continuously in the background at no extra cost.
From Reddit/HN/YC
- [Hacker News] 37signals publishes its Manager Playbook — the internal guide behind how Basecamp and HEY run teams.
- [Hacker News] Show HN: an 8-player survival Pong arena that loses a wall every time someone dies.
- [Hacker News] Forbes makes the case that smartphone innovation is moving beyond annual hardware upgrades.
- [Hacker News] Watchtowr drops a writeup on a Citrix NetScaler pre-auth RCE — CVE-2026-8452 is as bad as it sounds.
- [Hacker News] The case for phones having a guest lock feature — let someone borrow your phone without handing over your life.
- [Hacker News] Nicholas Rougeux details the painstaking making of the Four Books of Architecture visualization.
- [Hacker News] Miser is an open-source AI gateway built to route requests to the cheapest model that meets your quality bar.
- [Hacker News] A look back at Gary Kildall's design of knowledge retrieval systems for the PC — decades ahead of its time.