Jump to a brief
Grouped by topic first; AM/PM are editions inside the topic.
# 📡 Hacker News Afternoon Briefing — April 11, 2026 at 3:30 PM
**Top 5 Stories (ranked by AI/Startup relevance)**
**#1. Small Models Also Found the Vulnerabilities That Mythos Found** — Score: 9/10
AISLE's research challenges Anthropic's narrative that its Mythos model represents a unique breakthrough in AI vulnerability detection. They demonstrate that small, cheap open-source models — including one with just 3.6 billion parameters at $0.11 per million tokens — can replicate much of the same cybersecurity analysis when given properly scoped code. The key takeaway: the moat in AI cybersecurity is the system and scaffolding, not the model itself, and capability scales jaggedly rather than smoothly with size.
🔗 https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier
**#2. How We Broke Top AI Agent Benchmarks: And What Comes Next** — Score: 9/10
UC Berkeley researchers built an automated agent that achieved near-perfect scores on eight major AI benchmarks — including SWE-bench, WebArena, and GAIA — without actually solving any tasks. The exploits relied on poor isolation between agent and evaluator, exposed reference answers, unsafe eval() usage, and flawed scoring logic. They propose an "Agent-Eval Checklist" and are building BenchJack, a vulnerability scanner for evaluation pipelines, to prevent inflated capability claims.
🔗 https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/
**#3. 447 TB/cm² at Zero Retention Energy – Atomic-Scale Memory on Fluorographane** — Score: 4/10
Researchers propose a memory technology using single-layer fluorographane where each fluorine atom's bistable covalent orientation stores a binary bit, achieving 447 terabytes per square centimeter — exceeding all existing storage technologies by five orders of magnitude. The data persists indefinitely without power, with thermal bit-flip rates calculated at roughly 10⁻⁶⁵ per second at room temperature. A scanning-probe prototype already demonstrates proof-of-concept as a functional non-volatile memory device.
🔗 https://zenodo.org/records/19513269
**#4. Apple Silicon and Virtual Machines: Beating the 2 VM Limit** — Score: 2/10
Apple Silicon enforces a two-concurrent-VM limit via a closed-source kernel variable in XNU. The author discovered that booting a development kernel collection with specific boot arguments overrides this quota, successfully running nine macOS VMs simultaneously on an M2 Pro MacBook Pro. A useful hack for developers needing parallel macOS test environments, but niche in scope.
🔗 https://khronokernel.com/macos/2023/08/08/AS-VM.html
**#5. Dark Castle** — Score: 1/10
A retrospective on the classic Macintosh gaming trilogy spanning 1986 to 2008, from Delta Tao's pioneering black-and-white original through Beyond Dark Castle and the 2008 color remake Return to Dark Castle. Pure nostalgia — no AI or startup relevance, but a charming piece of computing history that clearly resonated with the HN crowd.
🔗 https://darkcastle.co.uk/
---