Perplexity hit $450M ARR in one month. Anthropic's Mythos leaked — and it's only for cybersecurity. Microsoft open-sourced agent security. Shopify gave agents the keys to your store. Eight stories from the week agents got dangerous.
Perplexity's revenue spiked 50% in a single month after pivoting from search to agents — $450M ARR and climbing. Anthropic accidentally leaked its most powerful model ever, then turned the embarrassment into Project Glasswing, a cybersecurity coalition with every Big Tech name attached. Microsoft open-sourced the first toolkit to cover all 10 OWASP agentic risks. Shopify handed AI agents the keys to 4.6 million storefronts. And Revolut launched an AI banker for 13 million UK customers. Eight stories from the week agents got dangerous — and indispensable.
Perplexity's pivot from AI search to autonomous agents just produced the most dramatic revenue spike in startup history. ARR jumped 50% in a single month to $450 million, according to the Financial Times. The catalyst: "Computer," an agent platform launched February 25 that orchestrates 19 specialized models to complete complex tasks, paired with a usage-based pricing model that charges beyond a monthly allocation. The company now has 100 million MAU and tens of thousands of enterprise clients. Prior to the shift, ARR was $305 million. In two years, it went from $16 million to nearly half a billion.
The fastest revenue acceleration in AI history — and it happened because Perplexity stopped competing with Google on search and started competing with everyone on execution. Usage-based pricing for agents is the business model the industry was missing. Watch every search company pivot by Q3.
Anthropic Leaked Its Most Powerful Model.
A misconfigured content management system exposed 3,000 internal Anthropic files in late March, revealing Claude Mythos — a model the company described as a "step change" in capabilities with an estimated 10 trillion parameters. The irony was brutal: a model promising "unprecedented cybersecurity capabilities" was outed by an elementary security flaw. Anthropic's response was swift and strategic. On April 7, it launched Project Glasswing, deploying Mythos Preview exclusively for cybersecurity — finding thousands of previously unknown zero-day vulnerabilities across major systems. The partner list reads like a who's who: AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, NVIDIA, and Palo Alto Networks.
Turning a catastrophic leak into the most elite cybersecurity coalition in history is either brilliant crisis management or the plan all along. Either way, Mythos finding zero-days before attackers do could redefine defensive security — if Anthropic can keep its own house in order.
Microsoft Open-Sourced the Security Shield Every AI Agent Needs.
On April 2, Microsoft released the Agent Governance Toolkit under MIT license — the first open-source framework to address all 10 OWASP agentic AI risks with deterministic, sub-millisecond policy enforcement. The seven-package system ships in Python, TypeScript, Rust, Go, and .NET, with native hooks into LangChain, CrewAI, Google ADK, and Microsoft's own Agent Framework. The Agent OS component intercepts every agent action before execution at p99 latency below 0.1 milliseconds. The project includes 9,500+ tests, continuous fuzzing, and SLSA-compatible provenance. Microsoft plans to donate it to a foundation for community governance.
This is Microsoft's play to own the governance layer of the agentic stack. Open-sourcing it is the smartest move — if every agent framework adopts Microsoft's security primitives, Microsoft becomes the de facto standard for enterprise agent compliance. The OWASP coverage makes it hard to justify building your own.
Shopify Just Gave AI Agents the Keys to 4.6 Million Stores.
On April 9, Shopify launched the AI Toolkit — a free, open-source plugin that connects Claude Code, OpenAI Codex, Cursor, Gemini CLI, and VS Code directly to the Shopify platform. Merchants can now type "update all product descriptions for summer" in Claude Code and watch it execute on their live store. The toolkit bundles 16 skill files covering every part of the platform, from inventory management to order fulfillment. The catch: the shopify-admin-execution skill has no built-in undo. When the agent makes a change, it's done. Two commands in Claude Code, one click in Cursor.
Shopify just made every store on its platform programmable by any AI agent. The no-undo design is either brave or reckless — but it signals where commerce is heading. The merchants who figure out agent-driven operations first will have an insurmountable efficiency advantage over those still clicking through dashboards.
Revolut Launched an AI Banker for 13 Million Customers.
On April 9, Revolut began rolling out AIR — AI by Revolut — to its 13 million UK customers. The in-app assistant replaces multi-step navigation with a single conversational interface for spending analysis, investment tracking, subscription management, card freezing, and eSIM purchases. AIR operates under a zero data retention policy with third-party AI providers and can only access what the customer already sees. Starling Bank launched a comparable assistant in March; NatWest debuted one the same month. UK fintech is becoming the proving ground for agentic banking — where agents don't just advise, they act.
"AIR replaces multi-step app navigation with a single conversational interface covering the core functions of day-to-day banking."
The race to build the AI banker isn't about chatbots anymore — it's about agents that execute transactions. Revolut, Starling, and NatWest all shipping within weeks of each other means UK fintech just became the global testbed for agentic finance. The winner will be whoever earns trust fastest.
Gemini 3.1 Pro Leads 13 of 16 Major Benchmarks.
Google's Gemini 3.1 Pro, which rolled out globally in early April, now leads 13 of 16 major independent benchmarks — including SWE-bench Verified at 78.8%, GPQA Diamond at 94.3%, and ARC-AGI-2 at 77.1% (double the reasoning performance of 3 Pro). It ties GPT-5.4 Pro on the Artificial Analysis Intelligence Index at roughly one-third the API cost. The model supports a 1M-token context window, processes 8.4 hours of audio or 900-page PDFs in a single prompt, and scores 85.9% on BrowseComp for autonomous web research. Gemini has crossed 750 million users.