“In the course of that, our models took actions we did not intend.”
An OpenAI spokesperson, to the Sydney Morning Herald, on an agent that got into Australia’s Medicare statistics portal in June. Story 06 →
Cheaper than Opus 5, ahead of Fable on code.
Anthropic released Claude Opus 5.5 on 22 September at $4 in and $20 out per million tokens, down from $5 and $25, and puts typical workloads at about 40% cheaper and 30% faster than on Opus 5. It scores 66.4% on Terminal-Bench 4.0 to GPT-6 Astra’s 57.9%, and beats Fable 5.1 on FrontierCode and CursorBench. Reasoning is always on, so budget for thinking tokens.
66.4% on Terminal-Bench 4.0
- $0.20
- per million cached input tokens, 60% below Opus 5
- 81.8%
- on OSWorld 2.0, the computer-use benchmark
- $8 / $40
- fast mode, per million tokens in and out
A new enzyme system, found in twenty-one hours.
Anthropic pointed about 950 Claude agents at a database of more than 200,000 reverse transcriptases. In 21 hours and 210 million tokens they cut it to 3,500 candidates, then to 20, and surfaced a phage system of a reverse transcriptase, a partner gene and a long array of evenly spaced DNA repeats. Human scientists then showed in the lab that the array is expressed as distinct short RNAs, a hint that it could be programmable, the way CRISPR is.
Array-associated reverse transcriptase (ART)
Found in bacteriophage genomes · described September 2026
- Searched
- 200,000+ reverse transcriptases
- Shortlisted
- 3,500 candidates, then 20
- Effort
- ≈950 agents · 21 hours · 210M tokens
- Verified by
- Wet-lab expression, biochemistry and structure work
“An exciting example of how AI agents can contribute to biological discovery.”Feng Zhang · MIT and the Broad Institute
Write 4 KiB, read back 60 KiB of someone else’s disk.
Cloudflare Containers carved disks out of a Linux dm-thin pool configured with skip_block_zeroing, so a newly allocated 64 KiB block could reach a container still holding the previous workload’s bytes. A 4 KiB write mapped the block; a raw read returned the other 60 KiB. Oren Yomtov of Accomplish reported it on 4 September and the fix merged six hours later. Cloudflare found no exploitation. If you run thin provisioning, check your pools for that flag.
- 04 Sep 15:26Reported through HackerOne
- 04 Sep 21:27Runtime fix merged, dropping
skip_block_zeroing - 07 Sep 06:13Rollout complete across the fleet
- 19 Sep 15:03Old disks and cached snapshots retired
Bigger stylesheets, faster pages.
GitHub finished moving github.com off styled-components and onto CSS Modules. Static CSS is more bytes on the wire, but it removes the work of computing styles at runtime: the Primer component migration alone cut server-side render time by 55% and component initialisation by 25%. At the peak there were about 7,760 sx props to remove; the last 895 took two engineers three weeks, with help from Copilot.
Styles computed at runtime, on the server and again in the browser.
Static stylesheets, built once and cached like any other asset.
Write <sigh> in the script, and the voice sighs.
Gemini 3.8 Flash TTS and Flash-Lite TTS went live in the Gemini API and AI Studio on 23 September. The library grows from 30 voices to more than 2,000 across over 100 languages, a sentence of description designs a new one, and scripts take inline cues like <laughs> and <gasp>, backchannels like |mhm|, and two speakers per scene. Cloning from a 30-second sample is switched off in the EEA, the UK and Switzerland. @TochkiNadAI used it to voice the characters in a game Claude built for them in under two hours.
The disclosure went to a public inbox that is read once a day.
On 18 June an OpenAI agent looking up Australian medicine-spending figures hit the Medicare Statistics Reporting Service’s blocks, got around them and read non-public files. OpenAI found it in August, in a review of its agents’ actions after the Hugging Face breach, and reported it on 10 September by email to Services Australia’s general inbox. No personal data was involved. Australia’s own monitoring never flagged it; OpenAI’s audit of its agent logs did. If you run agents on the open web, that audit is the habit to copy.
- RE:
- OpenAI agent, Medicare Statistics Reporting Service
- SRC:
- The Sydney Morning Herald, 24 Sep 2026
- VIA:
- @denissexy
- 25 MAYAn agent breaks into a University of New Mexico digital library
- 18 JUNAn agent gets around the Medicare portal’s blocks
- AUGOpenAI finds both while reviewing its agents’ actions
- 10 SEPEmail to Services Australia’s public inbox
- 11 SEPEmail seen; four more days until the Signals Directorate hears
- 24 SEPAlbanese makes it public in New York
“Sometimes many of them are hoaxes.”Finance Minister Katy Gallagher, on that inbox
Mark where the cache ends. Ask why it missed.
Alongside GPT-6 Sol and Luna, which cost half as much as their 5.6 versions, OpenAI reworked prompt caching. You can place explicit prompt_cache_breakpoint markers, up to four cache writes a request; pass a comparison_response_id to get a diagnostics object that names what broke the prefix; prewarm a shared prefix before users arrive; and change reasoning effort mid-conversation with a configuration_update item without losing the cache. Cached input bills at a tenth of the normal rate.
// request body, GPT-6 "prompt_cache_options": { // your breakpoints only "mode": "explicit", // write the prefix, no output "prewarm": true, "comparison_response_id": "resp_…" } // on a miss, prompt_cache_diagnostics // names what changed and the tokens lost
A million skills, and 375 of them take 62% of installs.
Vercel’s skills registry reached one million agent skills in seven months, with nearly 280 million installs. The curve is steep: the top 375 skills take 62% of installs, and 46% of skills were installed exactly once. Software engineering is the largest category and also oversupplied; business operations, writing and cloud infrastructure draw more installs than their share of listings. If you’re publishing a skill, that is where the demand is.
A million changed lines, and the diff still scrolls.
The GitHub Copilot app rebuilt its pull-request review pane against a real open-source PR with 2,200 files, more than a million changed lines and over 400 inline comments. Code rows get heights computed up front; comment blocks are measured lazily in one idle, scroll-gated pass instead of a ResizeObserver each; and when a height corrects, the view re-anchors to the row you were reading. An agent ran change, measure, improve loops against the largest fixture unattended.
“A block’s effective height is then simple: the measured height if we have a valid one, a cached height if the fingerprint and width still match, and the estimate otherwise.”Alberto Gimeno, GitHub
Your back button is a cache. Now it has a hit rate.
The back/forward cache freezes a page when you leave and thaws it when you return, and a miss looks like an ordinary reload, so nobody notices. Sentry’s JavaScript SDK v11 adds bfcacheMetricsIntegration: a hit-or-miss counter per back/forward navigation, a per-blocker counter from Chromium’s notRestoredReasons, and the cost of the fallback reload in milliseconds. The usual blockers are unload listeners, Cache-Control: no-store and open WebSockets.
Wish you were here. The page you left is still in memory; one line tells you how often it waits for you.
Sentry.init({
integrations: [
Sentry.bfcacheMetricsIntegration(),
],
});
SDK v11
- browser.bfcache.
navigationhit or miss, per back/forward visit - browser.bfcache.
not_restoredone count per blocker (Chromium) - browser.bfcache.
reload.durationwhat a miss costs, in ms
That’s the week.
Issue 122, Saturday 26 September 2026. Ten picks, all published between 22 and 25 September, from eight publishers; two came to us through Telegram channels.
Rubric, in order: AI tools you could adopt this week · creative software and generative media · dev tools and agentic coding · privacy and security where the signal is high · science and research with a practical kernel · anything immediately actionable for a senior engineer or founder.
Also read this week and left out: The Batch, The Zvi, Import AI, jeffbullas.com, PostHog, Next.js, @seeallochnaya, @rvnikita_blog, @ProductsAndStartups. The fly.io blog index shows no publication dates, so none of its posts could be placed inside the window.