Saturday, May 9, 2026

Good Saturday, NOLA. May 9th is about understanding the real impact of AI on work and security—not the hype, but the hard lessons. AI is breaking vulnerability disclosure norms, Claude is learning to explain its reasoning, and someone built Git for AI agents. Plus: the quiet story that Anthropic is hiring 10x faster than the industry is laying off.

AI Safety & Security: The Hard Problems

AI is breaking two vulnerability cultures

Security researchers rely on coordinated disclosure—quietly reporting bugs before going public. AI makes this harder: you can't keep an exploit secret when the vulnerability exists in a model weights file that's already been shared globally. This piece cuts through the noise to explain why traditional vulnerability disclosure is collapsing and what we do about it. Discussion on HN.
Hacker News

Claude Code CVE-2026-39861: Sandbox escape via symlink

A real vulnerability found in Claude Code's sandbox—agents could escape using symlinks. It's now patched, but this is the kind of thing you want to see caught and fixed early. The fact that this exists and is being openly tracked says something good about how the space is maturing.
GitHub Security Advisories

Can LLMs model real-world systems in TLA+?

Formal verification—proving your system is correct before it breaks—is hard. Can AI help? This deep-dive explores whether LLMs can write TLA+ specifications (a formal logic language) for real systems. Spoiler: it's complicated, but there's promise. HN discussion.
Hacker News

Understanding Claude's Thinking

Teaching Claude Why: Explaining reasoning in language

Anthropic released a new research piece on how Claude learns to explain its own reasoning—why it chose a particular answer, not just what the answer is. This matters for trust and debugging. When Claude can tell you *why* it did something, you can actually use it for high-stakes work. HN discussion.
Anthropic Research

A recent experience with ChatGPT 5.5 Pro

Real-world test drive: a mathematician uses GPT-5.5 Pro on actual work problems. It's honest about what the model can and can't do—useful for getting a sense of where the ceiling is right now. Not hype, just evidence. HN thread.
Hacker News

Tools & Things People Built

Git for AI Agents (re_gent)

Someone built version control for AI agent workflows—tracking agent state, branching, and rollback like you would for code commits. This is exactly the kind of boring infrastructure that makes agents actually usable in production. Show HN discussion.
Hacker News

Using Claude Code: The Unreasonable Effectiveness of HTML

Thariq Shihipar from the Claude Code team argues that HTML—not Markdown—is the right format for agents to work with. It's semantic, it's flexible, and it lets agents actually see structure. A smart piece on what makes agentic coding actually work in practice.
Simon Willison

Ben's Builds #3: An Email App

Ben Tossell is shipping an AI-native email app. This is the kind of "I'm actually building something" energy the industry needs—not a white paper, not a panel discussion, just shipping. Watch what works and what breaks.
Ben's Bites

CyberSecQwen-4B: Small, specialized security models

A 4B parameter open-source model trained for cybersecurity work—fast enough to run locally, specialized enough to actually be useful. Part of a broader shift: instead of one huge model for everything, people are training lean specialists.
Hugging Face Blog

The Quiet Structural Story

Anthropic growing 10x/year while the industry lays off >10%

While OpenAI and other labs are cutting costs, Anthropic is hiring aggressively—10x headcount growth. This is a structural signal about who investors think will win. It's quiet but important context for understanding where the industry is actually moving.
Latent Space

Cloudflare says AI made 1,100 jobs obsolete—while revenue hits record high

CEO Matthew Prince: because of AI, we don't need as many support roles. The company is cutting headcount while growing revenue. This is the actual economic story—not "AI will kill jobs someday," but "it's happening now, and businesses are optimizing."
TechCrunch

Worth a Listen

The Week the AI Story Shifted

A week-in-review podcast that steps back from the noise. The narrative around AI is forking: away from job-apocalypse panic toward a more realistic picture of how AI actually diffuses through the economy. Worth an hour of your time if you're thinking strategically.
AI Daily Brief Podcast

Also

Today’s Sources