Progress used to be limited by how quickly we could find the next bug. Now it's limited by how quickly we can patch it.
Anthropic · Project Glasswing · May 2026
Ephemeris.
Finding bugs is now the easy part.
Project Glasswing put Claude Mythos against a thousand open-source projects and forty-eight partner products for one month. The model returned ten thousand high or critical findings. Mozilla alone got two hundred seventy-one new Firefox bugs. The disclosure queue is now the bottleneck — average patch time runs two weeks per bug, and seventy-five of five hundred thirty reported flaws have shipped a fix.
Claude could do 94% of the job. It does 33%.
Anthropic's new "observed exposure" measure crosses theoretical capability with what Claude.ai is actually used for. The gap is wide everywhere, but widest in Computer & Math: theoretical coverage runs to 94% of tasks, real-world usage to a third. The bottleneck isn't whether the model can do it — it's trust, software integration, and verification workflows that haven't been built.
Forty-three percent of work is not on the bench.
Carnegie Mellon and Stanford mapped today's leading agent benchmarks to O*NET's labor-statistics taxonomy. The benchmarks cover 56.5% of US work activities; the missing 43.5% is mostly the economy. Software-engineering tasks dominate by an order of magnitude, while administrative support and management — the largest categories of US employment — barely register.
arXiv 2603.01203 · mapped to BLS occupational data
The agent shipped your app. without seams.
A coding agent writes a Next.js front end, a Supabase Edge Function, a Postgres query, and ships before lunch. None of it has tracing. Sentry's piece walks through wiring a single trace through all three layers so the N+1 query the agent quietly introduced becomes visible. The trick that makes the AI assistant useful afterwards is the MCP server — not its training data.
// One init wires browser, edge function, and Postgres into one trace. $ Sentry.init({ dsn: process.env.SENTRY_DSN, tracesSampleRate: 0.1, integrations: [ Sentry.supabaseIntegration(supabaseClient, Sentry, { tracing: true, breadcrumbs: true, }), ], }); // Now the agent's hidden N+1 shows up under perf → slow transactions. $ sentry mcp install --client cursor # for the *next* fix » "Use the MCP, not the training data."
Half a million requests
per second.
Bun 1.3.14 ships a built-in image pipeline (seventy times faster than sharp for metadata), experimental HTTP/3 in Bun.serve(), multiplexed HTTP/2 in fetch, and a global virtual store that takes a warm install from 841 ms to 115 ms on macOS. HTTP/3 isn't production yet. Everything else is.
An AI agent gets underwritten for the first time.
ElevenLabs paired its voice-agent platform with the new Artificial Intelligence Underwriting Company to issue the first AIUC-1 insurance certification. The model: like SOC 2, but for what happens when an agent hallucinates a refund, leaks a transcript, or takes an unauthorized action. Procurement now has a third party to point at.
Certifying ElevenLabs Agents against hallucination, unauthorised action, data leakage, and security failure across 14 risk categories and 5,835 individual tests.
?Fast, fine, three times the price.
Zvi works through Gemini 3.5 Flash. Verdict: a niche speed pick for agentic loops, not a daily driver. The Flash price tripled at the same time — fifty cents to a dollar fifty on input, three to nine on output — and on Artificial Analysis's index it lands behind both Opus and GPT-5.5. "If flash 3.5 had stayed at $0.5 it would be an insanely insanely exciting release."
Ephemeris.
Saturday Edition · Vol. I, No. 35
Today's read
Issue 035 followed a single thread: the bottleneck has moved. Models can find more bugs than maintainers can patch, do more of the job than companies have deployed them to do, and pass more benchmarks than reflect the work that actually pays. Three pieces about the gap, three about the tools narrowing it, one about a model that costs three times more to do the same thing.
Sources today
Anthropic Research (Project Glasswing · labor market impacts), The Batch (issue 354, on CMU and Stanford's benchmark-to-O*NET mapping), Sentry Engineering, Bun, ElevenLabs, Don't Worry About the Vase.
Rubric
AI tools you can adopt this week. Dev tooling and agentic coding. Privacy and security with high signal. Research with a practical kernel. Anything immediately useful for a senior engineer or founder. No marketing, no opinion-only.