Benchmark pits frontier LLMs against fresh real-world vulns

TLDR: A new live vuln benchmark tests if frontier LLMs can really find bugs, while Microsoft and AGIBOT lean harder into agentic assistants on screens and in robots.

N-Day-Bench launches live vuln benchmark for LLM code auditors

N-Day-Bench is a new evaluation that tests whether frontier large language models can find known security vulnerabilities in real open source repositories as of 2026-04-14. Each month it pulls fresh cases from GitHub security advisories, checks out the repository at the last commit before the patch, and gives models a sandboxed bash shell to inspect and execute code.

Static vulnerability discovery benchmarks age quickly because the vulnerabilities leak into model training data and scores drift toward measuring memorization. N-Day-Bench matters if you are experimenting with AI code reviewers, secure-by-default agents, or automated patch bots, since it focuses on realistic bug hunting rather than synthetic patterns. The monthly refresh aims to keep the test set ahead of training contamination, though there is still no public leaderboard or standardized protocol yet.

If you are building security-focused agents, this is worth tracking or even integrating into your own eval suite to compare tools against a shared, evolving target.

Read more →

---

Microsoft pushes Copilot toward always-on agentic workflows

CNET reports that Microsoft Copilot is being reoriented toward the "agentic AI" model as of 2026-04-14, with inspiration from OpenClaw and its descendants. Nvidia has already shipped its NemoClaw reference stack with safety guardrails such as full action logging, and Anthropic now lets some Claude subscribers run longer lived, task-completing agents.

For application and infra teams betting on agents, the signal is that Microsoft is not just adding more chat modes. The company is testing always-on Copilot style assistants that can own multi step tasks from start to finish, with lifecycle, monitoring, and permissioning closer to real services than chatbots. That means you should expect APIs, policy controls, and deployment knobs that look more like workflow engines than UX helpers.

Microsoft is expected to reveal more at Microsoft Build 2026, so if you are in the Windows, Microsoft 365, or Azure ecosystems, this likely affects how you expose your products to users and to Copilot itself.

Read more →

---

AGIBOT unveils Genie Studio Agent no-code robotics platform

AGIBOT has announced Genie Studio Agent, a zero-code application platform for building and deploying robot behaviors, targeting the gap between advanced embodied AI research and real world rollouts. The product is pitched at teams that want to configure robot tasks through high level interfaces instead of custom ROS nodes and bespoke control stacks.

For robotics engineers and integrators, this reflects the same shift software teams are seeing in LLM agents: orchestration, safety, and deployment are now the main friction points, not just perception and planning models. If Genie Studio Agent can make it practical for non specialists to define workflows, constraints, and environment assumptions, it could broaden who can field robots in logistics, retail, and light manufacturing, though hard real time and edge deployment details are still unclear.

If you are building agent stacks that eventually need to control physical systems, it is worth watching how Genie Studio Agent models state, recovery from failure, and human in the loop overrides.

Read more →

---

Quick Hits

• Microsoft is testing the deployment of always-on AI agents like OpenClaw. Omar Shaheen at Microsoft confirmed that new Copilot agents are meant to take tasks from start to finish, with more details likely at Microsoft Build in June 2026.

• Show HN: Ithihāsas – a character explorer for Hindu epics, built in a few hours A small app built with Claude CLI lets you navigate Mahabharata and Ramayana characters by relationships instead of linear text, a nice example of lightweight agentic UX on existing corpora.

• How to build effective reward functions with AWS Lambda for Amazon Nova model customization AWS walks through using AWS Lambda to run scalable reward functions for Amazon Nova tuning, covering Reinforcement Learning via Verifiable Rewards and Reinforcement Learning via AI Feedback plus monitoring in Amazon CloudWatch.

• Show HN: Mcptube – Karpathy's LLM Wiki idea applied to YouTube videos Mcptube (34 stars as of 2026-04-14) is an MCP server that indexes YouTube transcripts for semantic search and Q&A across long form lectures, useful if your agents need precise citations into video content.

• Enterprises power agentic workflows in Cloudflare Agent Cloud with OpenAI Cloudflare Agent Cloud now integrates OpenAI GPT 5.4 and Codex so enterprises can build and run AI agents closer to their network edge, with Cloudflare handling isolation and routing.

• Exploring the new servo crate Simon Willison digs into the new Servo Rust crate, which exposes an embeddable browser engine API suitable for headless web automation, potentially interesting as a low level substrate for browsing agents.

• Steve Yegge Steve Yegge relays a conversation claiming Google engineering's AI adoption looks similar to a heavy industry company, a reminder that internal tooling and culture lag far behind the marketing surface.