Wednesday, August 19, 2026Vol. I · No. 001Washington, D.C. · 6:00 AM
The finite technology newspaper
First Fold
Printed for the screen. Finished by breakfast.
Today’s edition
Three developments. One quiet desk. Your morning, finished.
One consequential development from each desk—when the evidence clears the bar.
Today’s paper takes about six minutes from front to back.
First FoldAI & Models · Page 2
Material update · Medium-high confidence
OpenAI rewrites its safety rules as Astra tests the old ones
The lab paused some frontier training while it revisits a framework written before models approached its highest cyber threshold.
3 min read · Checked 5:42 AM ET
What happened
OpenAI said preliminary evaluations of its upcoming Astra model were strong enough that it could not rule out the highest cyber-capability tier in its Preparedness Framework. On August 18, Axios reported that the company was rewriting that framework, had paused two weeks of deployment-focused reinforcement learning, and was keeping its largest planned frontier training run on hold.
Why it matters
This is a rare example of an internal safety threshold changing day-to-day model development. It also exposes a governance problem: capability tests are moving faster than the rules meant to govern them. Astra’s assessed level remains preliminary and comes from the company’s own evaluations, so the pause is more certain than the capability claim.
What to watch
Look for the revised framework, independent evaluations, release restrictions, and a clear account of what controls must be in place before large training runs resume.
First FoldWork & Tools · Page 3
New development · High confidence
Microsoft folds Deep Research into a paid workflow
Deep Research is leaving the consumer Copilot app, while Microsoft 365 Premium subscribers keep an in-depth research path through Researcher.
2 min read · Checked 5:44 AM ET
What happened
Microsoft began retiring Deep Research in the consumer Copilot app on August 18. Premium subscribers can continue producing detailed reports with Researcher. Saved work is not being deleted: Premium users can find it through Researcher, while Personal and Family subscribers retain access through chat history. Reports can also be opened and saved in Word.
Why it matters
AI work features are still fluid, and a tool’s name, location, or subscription tier can change while teams are building habits around it. The practical risk is less about lost files today and more about undocumented workflows that depend on a product surface the vendor can consolidate.
What to do
Update internal guides that point to Deep Research. Export important reports to Word, record the prompts and review steps behind recurring research, and confirm which users actually have Premium access before standardizing on Researcher.
First FoldSecurity & Privacy · Page 4
Material update · High confidence
France’s tax breach reaches 678,000 people and businesses
Officials began notifying victims as scrutiny turned to why the theft was not detected when the attacker’s access was blocked.
3 min read · Checked 5:47 AM ET
What happened
France’s tax authority, DGFiP, confirmed that a breach affected data tied to 678,000 individuals and businesses. The attacker’s access was blocked at the end of June, but the theft was not identified then. Its scale became clear after a public claim in August. The government convened a crisis meeting, ordered an audit overseen by the national cybersecurity agency, and began victim notifications.
Why it matters
Tax and identity details can make fraud attempts unusually convincing. The incident also highlights a familiar defensive gap: stopping access does not prove that defenders know what was viewed or removed. A later detection can leave victims exposed before notifications begin.
What to do
Reach tax services through a known official address rather than inbound links. Treat tax-themed calls, payment demands, and document requests with suspicion. Security teams should review privileged access, exfiltration telemetry, third-party pathways, and whether notification plans can operate before every forensic question is settled.
First FoldPlatforms & Power · Page 5
The quiet desk
Nothing cleared the bar today.
We reviewed the desk, but no development inside this edition’s reporting window was consequential, well-sourced, and distinct enough to earn the page.
“We would rather leave a page quiet than print a weak story.”
Candidates reviewed11
Minimum score70
Stories forced0
First FoldBack Page · 6
End of edition
You’re caught up.
Three stories. One honest blank. Nothing else is waiting.
The signals that could materially change these stories next.
OpenAI’s revised Preparedness FrameworkThe revised rules and final Astra cyber classification have not been published.Meaningful next signal: A public framework revision, independent evaluation, or documented resumption of paused workloads.
France’s DGFiP security auditThe root access failures, exfiltration path, and remediation plan are not yet public.Meaningful next signal: The ANSSI-overseen findings and operational measures promised for September.
Next First Fold lands at
6:00 AMWashington, D.C. time · weekdays during the pilot
The promise
At most one story per desk
Primary evidence whenever possible
No sponsored placement
Corrections appended, never hidden
Use the Left and Right Arrow keys to turn pages. Swipe works, too.