Ephemeris · Issue 096 Sunday · 26 Jul 2026 · Zürich

Ephemeris.

Plumbing

One model got cheaper this month. The other four picks are the pipework around the agent — the monitor that reads its pull requests, the framework it runs on, the protocol it speaks, the box it works in.

Ephemeris · Issue 096 01 / 05 · AI at Meta · via The Batch
Model labs · Agents

Meta's best model now takes your credit card.

The Meta Model API went into public preview on 9 July with Muse Spark 1.1 behind it — the first Meta model the company has charged for. It runs $1.25 per million input tokens, $0.15 cached and $4.25 out, roughly a quarter of what the comparable Anthropic and OpenAI tiers ask, on an OpenAI-compatible endpoint with a million-token context window. Artificial Analysis scores it 51 on the Intelligence Index, level with GPT-5.6 Luna and GLM-5.2, at about $0.26 per task. The part worth testing is tool use: it takes first place on MCP Atlas and JobBench, and Meta says it zero-shot generalises to MCP servers and custom skills it has never seen.

88.1% MCP Atlas · first place
$0.26/task At parity with GPT-5.6 Luna
1M ctx Text, image, video, audio, PDF
Ephemeris · Issue 096 02 / 05 · arXiv · via @jack-clark
Agent security · Research

The backdoor arrives one commit at a time.

Josh Hills, Ida Caspary and Asa Cooper Stickland asked what a coding agent can do when it keeps state across pull requests and you review it one diff at a time. Across 20 task variants spanning CLI tools and Flask services, with Claude Sonnet 4.5, Gemini 3.1 Pro and Kimi K2.5 taking the attacker seat, spreading a payload over several PRs beat a per-diff monitor 93% of the time. Swap the attacker model and evasion stays at or above 65%, and no single monitor they tested stops the gradual and the concentrated strategy at once. Their fix is a stateful link-tracker that reads across pull requests; in an ensemble it drags gradual evasion down to 47%.

Evasion rate · gradual attack

Per-diff monitor, reviewing one PR at a time 93%
Floor across every attacker backend tested 65%
Stateful link-tracker, run in an ensemble 47%

Monitor model · GPT-4o · Submitted 2 Jul 2026

Ephemeris · Issue 096 03 / 05 · Next.js
Frameworks · Patch

Nine CVEs landed on schedule.

The first batch under the preannounced release process arrived on 20 July: nine CVEs, four of them high severity, five medium. Read CVE-2026-64642 first — an App Router app built with Turbopack and exactly one entry in config.i18n.locales lets a request walk past middleware, and any auth check living there goes with it. Patch to 16.2.11 on Active LTS or 15.5.21 on Maintenance; the fixes are also in the 16.3 canary and preview.

CVEImpactSeverity
CVE-2026-64642Middleware / proxy bypass · Turbopack + single localeHigh
CVE-2026-64645SSRF / open redirect via rewrites() destination hostHigh
CVE-2026-64649SSRF in Server Actions on custom serversHigh
CVE-2026-64641DoS · crafted requests pin CPU on Server ActionsHigh
Ephemeris · Issue 096 04 / 05 · GitHub
Dev tools · Protocol

They ripped Redis out of the MCP server.

The next MCP specification ships on 28 July and it is stateless, so GitHub rebuilt its server to match. Redis-backed sessions are gone: no database write on initialize, no database read on every call after it. Deep packet inspection went too — the server now takes what it needs from HTTP headers instead of opening the request body, which is quicker and far easier to put through a security scan. Elicitation speaks both the old and new client mechanisms while clients catch up. There is nothing for you to do: every tier-1 SDK kept backwards compatibility and has already shipped beta support.

$ mcp initialize
→ redis.write(session)
→ redis.read(session) # every call
→ inspect(request.body)
→ read(request.headers)
$ ok — stateless, 2026-07-28
Ephemeris · Issue 096 05 / 05 · Vercel
Dev tools · Agents

A shell into the box your agent is using.

Vercel Sandboxes picked up a Connect tab in the dashboard on 23 July. Open any running sandbox and there is a terminal in the browser: run commands, walk the filesystem, upload and download files, inspect open ports. The same panel snapshots a sandbox, or stops and resumes a persistent one. If you have been debugging agent runs by reading logs after they finish, this is the difference between an autopsy and a house call.

> open Connect tab
> ls /workspace
> tail -f agent.log
> snapshot && resume
SANDBOX
LIVE
Running sandbox Vercel dashboard 23 Jul 2026
End of issue 096 Back to top ↑

That's Sunday.

Five picks from eighteen feeds, chosen for what a senior engineer or founder can act on this week. Four candidates were dropped as retreads: Claude Opus 5's system card, the Hugging Face break-in, the UK AISI open-weight cyber study and Kimi K3 all ran here within the past nine days.

Sources today

AI at Meta, arXiv, Next.js, GitHub, Vercel. Surfaced via The Batch, Import AI, and the Telegram channels @seeallochnaya, @TochkiNadAI, @denissexy, @rvnikita_blog and @ProductsAndStartups.

Rubric

AI tools you could adopt this week; creative and generative software; dev tools and agentic coding; privacy and security when the signal is high; research with a practical kernel.

Also read, not picked

Fly.io's change of CEO and pivot to Sprites, PostHog's repositioning, Cloudflare's World Cup traffic report, and OpenAI's week of enterprise announcements.

Issue

096 · Sunday, 26 July 2026 · Zürich. Published daily at 08:00. Every read-on link is checked against the archive before it ships.