Ephemeris · Issue 098 Tuesday · 28 Jul 2026 · Zürich

Ephemeris.

Tue 28 Jul2026

Today's issue is about grading the machines — what they find, what they can rebuild from the outside, and where you draw the line before they act.

Ephemeris · Issue 098 01 / 05 · Vercel
Security · Evals

231 real bugs. 25 models. One scoreboard.

Vercel rewound an open-source codebase to the commit before a batch of security fixes, picked 50 entry-point files, and had humans judge every finding into a golden set of 231. Scoring is a recall-weighted F2, so a vulnerability you miss costs more than a false alarm you raise. Then they pointed 25 models at it and published the bill next to the score.

Model · settingScoreCost
GPT-5.6 Sol · xhigh35.58$55.98
Grok 4.5 · high15.58$5.60
Fable 5declined the work—
Read on →
Ephemeris · Issue 098 02 / 05 · Import AI
Benchmarks · Agents

61,000 lines, rebuilt from the outside in.

Epoch and METR's MirrorCode hands a model a working program and takes everything else away: no source, no web, only the command line the binary exposes. The model has to infer the whole thing from its behaviour and write it again. 132 task instances span 25 programs and six languages, from a 16k-line phylogenetic tree parser up to Apple's 61k-line pkl.

$ mirrorcode --target pkl --probe cli-only
source access ...... denied
web access ......... denied
reference size ..... 61,000 loc

perfect score ...... 17 / 25 programs
above 99% .......... 4 more
never reached 100% . 8 / 25
hardest ............ ruff · giac_subset · mailauth
gotree rewrite ..... $100–400 per run
Read on →
Ephemeris · Issue 098 03 / 05 · Cloudflare
Privacy · Tooling

It's curl, for the protocols that hide you.

Testing Oblivious HTTP has meant parsing binary blobs by hand, cross-referencing three RFCs, and writing a throwaway script for every hop in the chain. Cloudflare Research put that behind one command and open-sourced it under Apache-2.0. pvcli speaks OHTTP, CONNECT proxying and MASQUE today, with Privacy Pass and post-quantum support on the list.

Fig. 1 — one OHTTP request, relay and gateway pvcli -vvv --ohttp --first-hop https://relay-cloudflare.ohttp.info \
  --proxy https://gateway.ohttp.info -X POST \
  --header "content-type: application/json" --data '{"test":1}' \
  https://target.ohttp.info/anything The same check by hand: encode the request, wrap it for the relay, hand-verify the gateway's response bit by bit against the spec. The verbose flag now prints each hop instead.
OHTTPCONNECTMASQUEPrivacy Pass · soon
Read on →
Ephemeris · Issue 098 04 / 05 · GitHub · via @TochkiNadAI
Agents · Creative Tools

Your agent has been reading the title. Now it watches.

Ask Claude about a video today and it answers from the title, the description, and whatever it can guess. The /watch skill closes that gap: it downloads the file, pulls frames with ffmpeg, gets a transcript, and hands both to the model in one pass. MIT-licensed, 11.3k stars, v0.2.0 shipped at the start of the month.

  1. 01Fetch. yt-dlp pulls from YouTube, TikTok, Vimeo, Instagram, X and 300-odd other sites, or reads a local .mp4 .mov .mkv .webm.
  2. 02See. ffmpeg extracts keyframes or scene changes, and the frame count scales to the runtime so a two-hour video doesn't eat the context window.
  3. 03Hear. Captions first because they're free; Whisper via Groq or OpenAI only when there are none to grab.
  4. 04Install. /plugin marketplace add bradautomates/claude-video, or npx skills add bradautomates/claude-video -g on other skill hosts.
Read on →
Ephemeris · Issue 098 05 / 05 · GitHub Changelog
Dev Tools · Governance

Let it label the issue. Make it ask before it closes one.

GitHub Issues now grades what its agents do to your tracker. Every supported action — labels, fields, type, assignees, close — arrives with a confidence rating and a written reason, and a repo admin picks the threshold where automatic becomes ask-first. In public preview across Agentic Workflows and the Copilot cloud agent.

01

Approvals

The agent suggests instead of applying. Changes wait in a panel on the issue itself.

02

Confidence

Each action is rated high, medium or low. High applies on its own; the rest queue for review.

03

Rationale

Every action records why it happened, whether it landed automatically or waited.

04

Scope

Set per repository. A two-person repo can run wide open while a busy public one holds anything below high.

GitHub is explicit that this is not a server-side boundary. It's a workflow convenience, so keep treating agent permissions as the real control.

Read on →
End of issue 098 Back to top ↑

That's today.

Five picks for Tuesday 28 July 2026. A security benchmark with the invoice attached, a reconstruction test, a privacy CLI, a skill that gives an agent eyes, and a dial for how much your tracker lets agents decide.

Sources today

Vercel · Import AI (Jack Clark) · Cloudflare · GitHub, surfaced via @TochkiNadAI · GitHub Changelog.

Also read

The Batch · Don't Worry About the Vase · OpenAI · Anthropic · Sentry · PostHog · Fly.io · Next.js · Jeff Bullas · @seeallochnaya · @denissexy · @rvnikita_blog · @ProductsAndStartups.

Rubric
  1. AI tools you could adopt this week
  2. Creative software and generative media
  3. Dev tools and agentic coding
  4. Privacy and security, when the signal is high
  5. Research with a practical kernel
  6. Anything a senior engineer or founder can act on
Rules

Nothing reprinted from a previous issue. Nothing older than thirty days. Three to five stories, never more, from at least three separate sources.