Get tomorrow's brief in your inbox
Today: Anthropic pulled OpenClaw from Claude subscriptions, forcing pay-as-you-go billing starting Saturday. AI agents now find real zero-day vulnerabilities faster than human researchers, flooding open source projects with reports. Zapier MCP lets Claude and ChatGPT control 9,000+ apps with two tasks per tool call.
A month ago, open source security teams were laughing at AI-generated vulnerability reports. They were obviously wrong, low quality slop. That changed suddenly in March. Linux kernel maintainers now field 5-10 real, valid security reports per day, up from 2-3 per week two years ago. The curl maintainer spends hours daily triaging. Multiple researchers are now filing duplicate reports for the same bugs because agents find them simultaneously.
Why the shift? Frontier models have hit a step function in capability. Thomas Ptacek's assessment: "Within the next few months, coding agents will drastically alter both the practice and the economics of exploit development." AI agents excel at vulnerability research because they encode correlations across vast codebases, pattern-match known bug classes (stale pointers, type confusion, allocator grooming), and never get bored searching.
For small businesses, this has two implications. First, if you maintain any open source code or tooling, expect an avalanche of security reports. Budget time to triage them because many are legitimate. Second, if you hire security consultants or pen testers, their pricing models are about to shift dramatically. Automated vulnerability discovery means assessments will get faster and cheaper, but also means you need to patch faster because attackers have the same tools.
Vulnerability Research Is Cooked | Greg Kroah-Hartman Quote | Willy Tarreau Quote | Daniel Stenberg Quote
Zapier's Model Context Protocol (MCP) implementation gives AI assistants access to 30,000+ actions across 9,000 apps without building custom integrations. Connect Claude or ChatGPT to Slack, Outlook, Gmail, Salesforce, and thousands more using natural language commands. No coding required for basic setup. Developers can connect to any MCP-compatible AI client (Cursor, Microsoft Copilot Studio, Mistral).
Cost: 2 Zapier tasks per tool call. Available on all plans.
This is useful if you already live in Zapier workflows. If you don't, Zapier Agents might be simpler (they're AI teammates trained to work across apps). For solopreneurs who hate repetitive tasks, this turns your AI into an actual assistant that can send emails, update CRMs, and create calendar events on command.
Zapier tested every AI social media tool they could find. The top picks: Buffer for tailoring posts to each channel, Publer for value pricing, FeedHive for content recycling, Predis.ai for full AI-generated content, Hootsuite for growing businesses, and Sprout Social for analytics and listening.
The tools use generative AI for post ideas and images, NLP for sentiment analysis, and machine learning for optimal posting times. If you're on the content hamster wheel, these cut scheduling time and help you react to trends faster. Pricing varies widely. Start with free tiers to test before committing.
Best AI Social Media Management Tools
Starting Saturday (April 4, 3pm ET), Claude subscriptions will no longer cover OpenClaw usage. You'll need pay-as-you-go billing billed separately from your subscription. Anthropic says third-party tools like OpenClaw create usage patterns their subscriptions weren't built for. Subscribers get a one-time credit equal to their monthly plan cost.
OpenClaw creator Peter Steinberger (now at OpenAI) tried negotiating but only managed a one-week delay. This follows a high-severity OpenClaw vulnerability (CVE-2026-33579, rated 8.1-9.8) that let anyone with pairing privileges escalate to admin and take over the instance. If you use OpenClaw for business, budget for API costs or switch to Anthropic's own tools like Claude Cowork.
Anthropic Blocks OpenClaw | OpenClaw Security Vulnerability
Zapier CEO Wade Foster's vision: the ideal team isn't 10 people with specialized skills. It's one domain expert who knows the customer and assembles a team of agents to tackle problems. Instead of 100 people split into 10 functional teams, you might have 50 generalists, each orchestrating agents for specialized tasks.
Foster breaks down workflows vs. agents simply. Workflows are deterministic (steps one through five in order). Agents are probabilistic (you give a goal and tools, the LLM decides how). Neither is better. Agents are less reliable and more costly. Run more things as workflows if you can. Use agents for prototyping and iteration. Once you've figured it out, lock it into a workflow.
The hiring shift: Zapier evaluates every candidate on AI fluency. They're looking for people who are curious, actively using tools, and "trying to live in the future." For people entering the workforce, this levels the playing field. Someone with 20 years of experience may have skills that are less relevant. Master tools like Zapier, Claude, or ChatGPT and you can rocket past them.
A startup founder used ChatGPT and Claude to write code, run customer service, and scale a telehealth business to significant valuation with no employees. Details are sparse in the article stub, but the premise is clear: AI coding assistants and chatbot customer service let one person do work that previously required a full team.
If you run a service business, this is your playbook. Use Claude or ChatGPT for coding (even basic web forms or automations). Use AI chatbots for tier-1 customer support. Focus your time on the parts that require domain expertise and relationships. The rest can be delegated to tools.
1-Employee Billion-Dollar Startup
Utah approved a one-year pilot letting Legion Health's AI chatbot renew psychiatric prescriptions without a doctor. Cost: $19/month subscription. The system can renew 15 lower-risk maintenance medications (Prozac, Zoloft, Wellbutrin, mirtazapine, hydroxyzine) for stable patients already on treatment. No new prescriptions, no controlled substances, no ADHD meds, no antipsychotics.
Patients must opt in, verify identity, prove existing prescription, and answer symptom questions. Red flags escalate to clinicians. State officials say it frees healthcare providers for complex cases and addresses shortages (500,000 Utah residents lack mental health access).
Psychiatrists are skeptical. University of Utah's Brent Kious thinks advantages are overstated and it won't increase access for those most in need (you'd already need a treatment plan to use it). If you're evaluating AI healthcare tools for your business, this shows where regulators are drawing lines: narrow scope, stable patients, low-risk medications only.
Chatbots Prescribing Psychiatric Drugs
OpenAI's COO Brad Lightcap is transitioning to "special projects" focused on complex deals and investments, reporting directly to Sam Altman. Denise Dresser (former Slack CEO, recently joined as chief revenue officer) takes over some commercial duties. CEO of AGI deployment Fidji Simo is taking medical leave for a neuroimmune condition. Greg Brockman will manage product while she's out. CMO Kate Rouch is stepping down to focus on cancer recovery.
The churn comes as OpenAI preps for IPO (possibly this year) while facing intensifying competition from Google, Anthropic, and Elon Musk's xAI. The company shut down Sora video app last month to focus compute on enterprise and coding tools.
OpenAI Executive Shuffle | OpenAI AGI Boss Leave
OpenAI acquired Technology Business Programming Network (TBPN), a daily tech talk show with less than 60,000 YouTube subscribers but high-profile guests like Altman, Satya Nadella, and Mark Zuckerberg. Sam Altman called it his "favorite tech show." Analysts are confused. The M&A strategy is tough to pin down (they also spent $6.4 billion on Jony Ive's device startup and have acquired software, cybersecurity, and health-tech startups).
One analyst's take: "They are kind of chasing vibes a little bit." OpenAI closed a $122 billion funding round, so they can afford to experiment. For small businesses, this is irrelevant noise unless you work in media or content strategy.
Anthropic purchased Coefficient Bio in a $400 million stock deal. The stealth biotech AI startup (launched 8 months ago) was using AI to make drug discovery more efficient. Founders Samuel Stanton and Nathan C. Frey previously worked at Genentech's Prescient Design. The team (around 10 people) joins Anthropic's health and life science team.
This follows Anthropic's October announcement of Claude for Life Sciences, a tool for scientific researchers. If you work in healthcare or biotech, Anthropic is positioning Claude as a research assistant. Expect more domain-specific tools.
Anthropic Buys Coefficient Bio
Anthropic filed documents to create AnthroPAC, which will make contributions to both parties during midterms. Funded by voluntary employee contributions capped at $5,000. This follows reports that AI companies contributed $185 million to midterm races. Anthropic also contributed at least $20 million to Public First, a Super PAC financing ad campaigns supporting regulatory agendas.
The political activity ramps up as Anthropic is enmeshed in a legal battle with the Defense Department over government use of Anthropic's AI models. For small businesses, this signals that AI regulation is coming and companies are fighting to shape it.
Moonbounce (founded by former Facebook business integrity lead Brett Levenson) raised $12 million (Amplify Partners and StepStone Group). The company provides real-time content moderation wherever content is generated (by users or AI). They've trained their own LLM to evaluate content against customer policy documents, respond in 300 milliseconds or less, and take action (slow distribution pending review, or block high-risk content immediately).
Customers include dating apps, AI companion platforms, and AI image generators. They're supporting 40 million daily reviews and 100 million daily active users. Tinder's head of trust and safety said LLM-powered services like this achieve 10x improvement in detection accuracy.
If you run a platform with user-generated content, this is the next generation of moderation. Previous approach: human reviewers with 30 seconds per piece of content, 50% accuracy, days after harm occurred. New approach: AI evaluation at runtime, integrated into the product.
A new lawsuit claims Perplexity's incognito mode is a "sham" and that the company shared user conversations with Meta and Google. Details are sparse in the article stub, but this is a reminder: free AI tools monetize through data. If you use AI chatbots for business, assume conversations are training data unless you're paying for enterprise plans with explicit privacy guarantees.
Perplexity Data Sharing Lawsuit
Two polls show public opposition to data centers. Harvard/MIT poll: 40% support, 32% oppose. Quinnipiac poll: 65% oppose AI data centers in their community, 24% support. Two-thirds worry about electricity price increases. More people would rather have an Amazon warehouse in their neighborhood.
Meanwhile, Microsoft, Google, and Meta are building massive natural gas power plants in Texas and Louisiana to power data centers (5-7.46 gigawatts). Turbine shortages mean 195% price increases by end of year and six-year delivery times. This is the AI infrastructure bet: natural gas will be necessary for success, AI won't need exponential power growth forever, and prices will stay stable.
For small businesses, this means cloud AI costs are likely to rise. If you're building AI-heavy products, budget for infrastructure price increases.
Natural Gas Data Centers | Public Opposition to Data Centers
The Axios supply chain attack used individually targeted social engineering to install malware on a maintainer's machine. The attacker cloned a real company founder's likeness, created a branded Slack workspace with fake employee profiles and OSS maintainers, scheduled a Microsoft Teams meeting, and during the call said something on the maintainer's system was out of date. Installing the "missing item" was actually a Remote Access Trojan that stole credentials.
If you maintain any open source code or manage developer teams, brief them on this attack. The time constraint of joining meetings late makes people click "yes" quickly. Verify software update prompts independently. Never install software during a meeting with people you haven't met in person or verified through trusted channels.
GitHub COO Kyle Daigle reports 1 billion commits in 2025. Now 275 million per week, on pace for 14 billion this year (if growth stays linear, which it won't). GitHub Actions grew from 500M minutes/week in 2023 to 1B in 2025, now 2.1B minutes this week. This is agents writing code. If you're hiring developers, expect their workflows to look different. If you're a developer, these numbers show how fast AI-assisted coding is scaling.
AI agents crossed a capability threshold in March. They're now finding real vulnerabilities, writing real code, and handling real business tasks at scale. The shift from "AI slop" to "real reports" happened suddenly, and open source maintainers are drowning. Expect the same pattern in other knowledge work domains: a sudden step function where AI output quality jumps from useless to legitimate.
The business implications are clear. If your competitive advantage relies on specialized knowledge (security assessments, code reviews, content creation), AI is compressing timelines and dropping prices. If you manage people, the hiring criteria are shifting toward AI fluency and domain expertise over specialized skills. And if you use AI tools for business, brace for price increases as infrastructure costs rise and platforms crack down on third-party access.
Watch for consolidation this quarter. OpenAI is buying everything, Anthropic is raising and spending, and smaller tools are getting squeezed out or acquired. Stick with platforms that have clear revenue models and enterprise commitments.