# Daily Summary for 2026-03-31 ## 2026-03-31 15:00:59 # AI Digest - March 31, 2026 ## Industry News - **Claude Code source leaked via npm**: Anthropic accidentally published a source map file exposing ~500k lines of TypeScript code for Claude Code, revealing internal architecture, hidden features like "Kairos" daemon mode, and upcoming multi-agent coordination. [link](https://x.com/Fried_rice/status/2038894956459290963) - **Axios npm supply chain attack**: Malicious versions 1.14.1 and 0.30.4 briefly shipped with a hidden RAT dependency affecting 300M+ weekly downloads, compromising major packages like auth0, Slack SDK, and AWS clients. [link](https://x.com/karpathy/status/2038949933859746264) - **Mercor AI breach (4TB leaked)**: Recruiting platform allegedly compromised by Lapsus group, exposing 939GB source code, candidate resumes, interview data, and internal VPN access across their $10B valuation startup. [link](https://x.com/AlvieriD/status/2038966789728420184) ## Tips & Techniques - **Ask agents "Are you missing context?"**: Simple but powerful prompt pattern from the Claude Code leak that helps agents identify gaps in their understanding before proceeding with tasks. [link](https://x.com/altryne/status/2038676458026189225) - **Agent harness matters more than model**: Analysis of leaked code shows Claude Code's performance comes from sophisticated harness engineering (retry logic, context management, tool definitions) rather than model alone—changing just the harness produces 6x performance gaps. [link](https://x.com/LiorOnAI/status/2038971482063569348) - **Debugging agent context with trace format**: Converting raw JSONL logs into structured views (not just feeding them back) cuts token consumption and produces more concise learned memory for agents. [link](https://x.com/Haofan_Wang/status/2038973883639099489) ## New Tools & Releases - **Pazi: AI DevOps team in Slack**: Launches with OpenClaw integration, executing tasks across Sentry/Linear/GitHub stack as an always-available team member responding to incidents and managing workflows. [link](https://x.com/ZvonimirSabljic/status/2038988794956075471) - **Merge Gateway for LLM routing**: Production-ready infrastructure for routing, fallback, cost guardrails and security across multiple LLM providers—solves the "stitched together stack" problem teams face. [link](https://x.com/shensi/status/2038966442894921741) - **Resend becomes agents' default email**: Claude Code now automatically selects Resend for email functionality without human intervention—first clear example of API-first products winning agent-driven distribution. [link](https://x.com/JasonToevs/status/2038992725622788259) ## Research & Papers - **Model harnesses create 6x gaps**: Stanford/MIT paper shows changing the harness around fixed LLMs produces massive performance variations—the same model can appear dramatically different based on prompt engineering, tool access, and context management. [link](https://x.com/omarsar0/status/2038971936382189793) - **FIPO: Fine-grained credit assignment for reasoning**: Qwen introduces Future-Influenced Policy Optimization using "FutureKL" to measure how current tokens affect entire reasoning trajectories, addressing the missing piece for scaling RL in reasoning. [link](https://x.com/sheriyuo/status/2038994381068542303) - **Seen2Scene: Training 3D on real partial scans**: Introduces visibility-guided flow matching enabling training on incomplete real-world data instead of synthetic scenes, advancing scan completion and text-to-3D generation. [link](https://x.com/angelaqdai/status/2038993071442936211) ## 2026-03-31 15:00:59 ## Security - **Google quantum breakthrough threatens crypto**: Two papers released showing significant progress toward breaking RSA-2048 and elliptic curve cryptography, prompting Google to move post-quantum transition deadline to 2029. [link](https://x.com/drakefjustin/status/2038973437886681415) - **Claude Code security is just a text prompt**: Leaked code reveals the "safety system" for dangerous cyber security work is literally a replaceable string—users can simply compile their own unrestricted version. [link](https://x.com/paoloanzn/status/2038944622039413224) --- *Curated from 400+ tweets across AI research, developer, and security communities* --- ## Emerging Trends ✨ **Claude Code Source Leak** (89 mentions) - NEW Claude Code's source code was accidentally leaked via a map file in their npm registry, revealing the entire codebase including safety systems, features, and implementation details. The leak has sparked widespread discussion about code security and open source. ✨ **Claude Code Cache Bugs** (34 mentions) - NEW A Redditor reverse-engineered Claude Code binaries and discovered two bugs causing cache invalidation issues: a string replacement bug in the custom Bun binary and a --resume flag that always breaks cache. This has been causing users to hit API quotas unexpectedly due to 10x-20x more expensive uncached tokens. ✨ **AI Agent GTM Strategy** (25 mentions) - NEW Discussion about how AI agents are now making purchasing decisions for developers, choosing tools like Resend, Stripe, and Firecrawl automatically. Companies need to optimize for agent discovery through API-first design, excellent documentation, and structured reputation rather than traditional marketing. ✨ **Pazi AI Team Launch** (18 mentions) - NEW Pazi launched as an AI team that works 24/7 in Slack, executing across entire tech stacks including Sentry, Linear, and GitHub. Powered by OpenClaw, it's now in beta and offering free credits to early users. 📊 **OpenClaw Adoption** (28 mentions) - CONTINUING OpenClaw continues to see strong adoption and integration into various tools and platforms, with mentions of new projects like Pazi (AI team in Slack), GitAgent (framework-agnostic agent standard), and various implementations. Discussion focuses on it being a foundation for agent-based tools.