Top Stories

Project Glasswing: Securing Critical Software for the AI Era

969 points · anthropic.com

The runaway story of the day is Anthropic’s Project Glasswing, a collaborative cybersecurity initiative that gives organizations building critical infrastructure — think operating systems, browsers, and core libraries — access to Claude Mythos Preview for vulnerability hunting. The model has already uncovered thousands of zero-day vulnerabilities in major systems, including a 27-year-old flaw in OpenBSD and bugs deep in the Linux kernel. Anthropic is backing it with $100 million in model credits and $4 million for open-source security orgs. The 431-comment thread reflects both excitement and unease: the same AI capabilities that find vulnerabilities for defenders could eventually be weaponized by adversaries. Glasswing’s bet is that giving defenders a head start during this transition window is the smartest move available.

Show HN: Brutalist Concrete Laptop Stand

710 points · sam-burns.com

Sometimes the most beloved Show HN posts aren’t software at all. Sam Burns built a fully functional laptop stand from solid concrete, complete with USB charge ports, a three-pin plug socket, and deliberately weathered rebar for that abandoned-building aesthetic. He used ammonia to corrode copper elements and carefully incomplete concrete mixing to nail the brutalist look, then added a trailing plant in an integrated pot. It’s almost certainly the heaviest laptop stand ever made — he needs a trolley to move it. The 218-comment discussion is pure delight, mixing sincere admiration for the craftsmanship with jokes about ergonomic RSI from desk accessories that weigh more than the desk.

System Card: Claude Mythos Preview

558 points · anthropic.com

Paired with Glasswing, Anthropic released the system card for Claude Mythos Preview — the frontier model powering the security initiative. The 409-comment discussion digs into the model’s capabilities at code-level vulnerability analysis, its performance on cybersecurity benchmarks, and what safeguards are in place to prevent misuse. The HN community is particularly focused on the tension between transparency (publishing a system card) and security (revealing what the model can do). It’s a real-time case study in responsible disclosure at the frontier of AI capabilities.

GLM-5.1: Towards Long-Horizon Tasks

442 points · z.ai

Zhipu AI’s GLM-5.1 is making waves as a model specifically optimized for long-horizon agentic tasks — the kind where an AI needs to plan, execute, and adapt across many steps over extended timeframes. The 164-comment thread is debating whether this represents a meaningful architectural advance or incremental progress dressed up in marketing. Either way, the focus on sustained multi-step reasoning rather than single-turn benchmarks reflects where the frontier labs see the next battleground: not just how smart a model is in a snapshot, but how reliably it can carry out complex workflows from start to finish.

NASA Lunar Flyby: Artemis II’s Far Side Photos

437 points · nasa.gov

The Artemis II crew conducted a seven-hour pass over the Moon’s far side on April 6, capturing photographs of regions no human has ever seen at this resolution — including Ohm crater and Vavilov Crater in striking detail. The crew also photographed a 54-minute solar eclipse from deep space, with the Moon backlit by the Sun and Earth visible in the distance. The 112-comment thread is a mix of technical orbital mechanics and uncharacteristic HN sentimentality. After years of Artemis delays, having astronauts actually out there shooting photos of the far side is quietly electric.

Cloudflare Targets 2029 for Full Post-Quantum Security

289 points · blog.cloudflare.com

Cloudflare just drew a line in the sand: full post-quantum security, including authentication, by 2029. The urgency comes from recent breakthroughs — Google improved quantum algorithms for breaking elliptic curve crypto, and Oratomic estimates neutral atom computers could crack P-256 with just 10,000 qubits. Cloudflare already has post-quantum encryption on 65% of human traffic (defending against harvest-now-decrypt-later attacks), but authentication remains the critical gap. The 92-comment discussion highlights that migration isn’t just flipping a switch — it requires years of work across complex dependency chains, third-party vendors, and the need to fully disable legacy systems to prevent downgrade attacks.

S3 Files and the Changing Face of S3

218 points · allthingsdistributed.com

AWS is finally bridging the gap between object storage and filesystem semantics. S3 Files lets you mount S3 buckets as network filesystems on EC2, containers, or Lambda — addressing the persistent friction of tools that expect local file APIs but data that lives in S3. The clever bit is the “stage and commit” architecture: changes accumulate in EFS, then sync to S3 roughly every 60 seconds, preserving both NFS consistency and S3’s atomic guarantees. With lazy hydration, read bypass hitting 3 GB/s throughput, and bidirectional sync, this could reshape how ML pipelines, genomics tools, and agentic frameworks interact with cloud storage. The 65-comment thread is cautiously optimistic.

Google Open-Sources Scion: Agent Orchestration Testbed

163 points · infoq.com

Google released Scion, an experimental platform that functions as a “hypervisor for agents” — orchestrating Claude Code, Gemini CLI, Codex, and others as isolated, concurrent processes with dedicated workspaces and credentials. The design philosophy is isolation over constraints: instead of embedding behavioral rules in agents, Scion enforces boundaries at the infrastructure level through containerization, git worktrees, and network policies. Agents can run locally, on remote VMs, or in Kubernetes clusters. Google even shipped a demo game called “Relics of the Athenaeum” showcasing multi-agent collaboration. The 46-comment discussion is intrigued by the model-agnostic approach.