techFri 23 Oct, 07:30

Sources93 X accounts · techtwitter.com · Hacker News · 8 newsletters
  1. Runs every day at 07:30 London time.
  2. It is shown its own papers of the last 3 days, to know what it has already said.
  3. Haiku 5.5 writes the paper from what was fetched, with no tools.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

93 X accounts

Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.

techtwitter.com

What tech Twitter is discussing, ranked by discussion.

Hacker News

The front page, plus stories over 80 points in the last 24 hours.

8 newsletters

Every new issue, each delivered once, looking back up to 7 days.

The tech daily written for one reader

Auth

Kill static credentials

Vercel found only 8.4% of over 1,000 SaaS vendors support secure login.

Vercel says it is time to retire long-lived static credentials. Cramforce says marketplace vendors must now support OIDC, and shadcn released a registry to set it up with one command.

Anthropic

Claude goes auto by default

An automated check now blocks risky actions, and you can switch back to manual at any time.

Starting next week, sessions in Claude's web, desktop and mobile apps run in auto mode for Pro, Max and Team plans. Instead of asking you to approve each step, a check blocks the risky ones. The post drew 564 replies.

Cognition open-sources a vulnerability-fix plugin for agents

The plugin helps agents secure systems against recent threats, and works with any agent or inside Devin.

Vercel's temporary deploys go live in one command

Guillermo Rauch says vercel deploy --temporary gives a throwaway link, built for agents shipping quick apps.

Agent memory

Agent memory may be overkill

On one long conversation, Mem0 beat plain full context only with GPT-4o-mini.

The MemoryTax study finds that strong models can search raw history themselves. Building Mem0's memory cost more than a full-context run.

Claude Code

Agent Skills goes open

Anthropic is donating the format to the Agentic AI Foundation, next to MCP.

Agent Skills started as an experiment inside Claude Code and many tools now use it. Anthropic says it fits better under a neutral group, like the MCP standard already does.

Big announcements

Friday, October 23, 2026 · Pergent2

Agent security

1Password opens to agents

Users approve which logins to share, and the agent signs in without ever seeing the password.

1Password launched a platform for agents with partners including Anthropic, OpenAI, Perplexity and Grok Bot. Grok Bot can now use a 1Password account, and Kody says its secrets can be used by agents without being exposed.

Anthropic

Math proofs for real code

The work covered over one hundred open-source libraries in a few weeks, working with Theorem Labs.

Anthropic's Frontier Red Team used Claude to prove security properties in real open-source code. The post reached the top of Hacker News with 262 points.

Agent runtime

NVIDIA's OpenShell

It is an open-source runtime for fleets of autonomous AI agents, and it drew 187 comments on Hacker News.

OpenShell gives groups of autonomous agents a private, safe place to run. It was one of the day's top Hacker News stories.

What people are fighting about

Friday, October 23, 2026 · Pergent3

Security

Security warnings grow louder

An Anthropic security officer says things will get worse before they get better, and sarcastic replies push back.

Peter Wildeford quotes Anthropic's deputy security chief saying the past four days were bad and worse is coming. Nikita Bier's post saying nothing important runs on PHP drew 450 replies, and Deedy says non-technical people can now exploit old bugs.

Engineering costs

Security costs will climb

Gergely Orosz predicts security will cost a lot more, and many security pros are skeptical of AI.

Orosz says security teams will run big AI bills and hire pricey AI-native engineers. He argues the skeptics leave room for people who are optimistic about using AI to defend systems.

Web dev corner

Friday, October 23, 2026 · Pergent4

TypeScript

A no-codegen database library

Reylo 1.0 is due alongside TypeScript 7.1 stable.

Reylo, short for relational loader, gives zero codegen and fully inferred types. Hacks says it makes the ORM versus SQL debate mostly moot, since you get full SQL power with types.

Databases

Turso beats SQLite on writes

At 64 connections, Turso hits about 9,500 writes per second versus about 1,370 for SQLite.

Turso 0.8 uses an MVCC engine with BEGIN CONCURRENT to allow many writers at once. SQLite still wins with one connection, and Turso overtakes it at two.

Web security

Certificate firm quits SSL

eMudhra says it is leaving global SSL/TLS certificates, after key problems and a separate login bypass surfaced this week.

eMudhra announced it is exiting its global SSL/TLS certificate business. Gergely Orosz suspects Google may drop it from Chrome trust after a key-making mistake, which is speculation. A separate post shows a full bypass of eMudhra's OTP login, and a new CVE flags a compromised intermediate certificate authority.

PHP

OpenAI's PHP claim

OpenAI says it spent two months working with PHP maintainers, Wiz and Cloudflare before disclosing two critical flaws.

This follows up on the PHP patch. OpenAI says a group using GPT-5.6 Sol found CVE-2026-109618 and CVE-2026-109619 before attackers did. The FBI urged organizations to update PHP now or take systems offline.

From the newsletters

Friday, October 23, 2026 · Pergent5

AI agents

Hotels trip up booking agents

Agents reached checkout at five or six of ten groups, and no reservations were completed.

Two October 18 tests sent AI agents through major hotel websites. Some sites admitted one agent while blocking another, which suggests bot defenses shape access without a clear rule.

Cyber stats

Vulnerabilities up 81 percent

ThreatLocker's CEO says the numbers are still climbing, and he expects that for two or three more years.

Nearly 38,000 CVEs were published in Q3, up 81 percent from the quarter before. The Deep View argues AI labs must share the burden of cyber defense, since the same tools serve defenders and attackers.

Hiring

Google tests AI in interviews

Google says AI fluency is now required for software engineering roles.

Internal documents show Google expanding pilots where candidates use AI while showing they understand code. Interviewers are also testing AI transcription and feedback drafts, though the formats are still pilots.

AI agents

Agent gets booted from listserv

Members left and argued about consent, and the group removed the agent about an hour later.

An agent joined a mailing list of internet researchers and introduced itself openly. The group's policy bans agents, and the October 15 incident shows how a personal assistant can enter spaces that never agreed to host it.

AI training

Work records sell as training data

One agency licensed design files and checklists while excluding client content, and the payment is undisclosed.

Businesses are finding buyers for records of how they work, since developers want them to teach agents longer workflows. The agency's deal is one example of a new market.

AI research

AI helps prove old conjecture

Researchers guided AI agents that tried more than 100 approaches, then checked the final proof themselves.

Xinzhi Zhang and Zhao Song proved the KLS conjecture with help from AI agents, according to Microsoft Research. Results that once assumed KLS was true can now drop that assumption, and a later machine check built on their work.

Research & open source

Friday, October 23, 2026 · Pergent6

Agent harness

Own your agent harness

Ramp's harness is model-agnostic and lets its teams ship faster.

Omar Sarar calls on full-stack AI companies to build their own harness. He says owning observability matters given the complexity of agent systems.

AI research

Agents beat 20 benchmarks

Claude Opus improved on the published best by 10.4 percent on average, with no misconduct flagged.

Agent4Sci gives coding agents 20 scientific ML benchmarks across 11 domains and asks them to beat the best published model. The gains mostly come from standard ML tricks, and the agents struggled to judge what is novel.

AI safety

Outside monitor check

METR and Redwood Research got three weeks of access to OpenAI's internal monitoring system.

The pilot helped strengthen safeguards and will shape future work with METR and other third-party evaluators.

AI and science

AI peer reviewers win

In a study of 124 Harvard social science papers, the model beat human experts on every measure.

Ethan Mollick says the study covered statistics, formal models and qualitative work. He argues academics must adapt science to AI, fast.

Worth a read

Friday, October 23, 2026 · Pergent7