PyyanThe AI Index

Pyyan / News

What happened in AI

A dated log, newest first. Every item was read from its source before it was written up, and every one carries the link so you can go and disagree with it.

68 stories · 4 straight from the source · showing 1 to 10, page 1 of 7

The days

By kind

Who keeps appearing

19 September 2026

2d ago
PolicyWhite House

Trump says he is forming an AI Force, and did not say what it would do

No detailson the AI Force's powers, budget or appointee

Every other governance move this fortnight named a mechanism: an evaluator with employee level access, a kill switch study with a November deadline, a liability rule. This one named a brand.

On Saturday 19 September the President posted on Truth Social that he is forming the AI Force, much like I did Space Force, and will soon name an AI czar to run it, a job for which, he wrote, only high IQ individuals need apply. The framing was protective rather than regulatory: the administration will not, he wrote, hinder or stifle the growth of this incredible industry, and he cast the move as defending AI from Democrats who want to weaken it, while saying existing law would be used against harmful conduct. No structure, no budget, no statutory authority and no name were given.

Why this one is different

The comparison he chose is the useful part. Space Force was created by an act of Congress, in the 2020 defence authorisation, with a budget line and a chain of command. A post is not that. What makes this worth recording anyway is the direction: in a fortnight where the labs themselves asked to be slowed down and the Treasury Secretary refused them a liability shield, the President's contribution is an office whose stated purpose is to keep the industry from being slowed.

The labs asked for brakes. The White House announced a flag.

How we got here

  1. 3 Sep 2026Two members of Congress propose making superintelligent AI a crime.
  2. 12 Sep 2026The heads of Anthropic, OpenAI and xAI all call in public for the frontier to slow.
  3. 15 Sep 2026The Treasury Secretary refuses the labs a liability exemption.
  4. 18 Sep 2026California orders work on a kill switch and onsite auditors, with dated deadlines.
  5. 19 Sep 2026Trump announces an AI Force and an AI czar, with no details.

What it does and does not mean

Nothing exists yet. There is no order, no appointment, no budget and no described power, and until there is, the only AI governance in the United States with dates attached is a state executive order and a set of bills that have not moved. That is the thing to watch rather than the name. Federal policy has now been announced as protection of the industry at exactly the moment the industry's own leaders asked for restraint, which means the next fight is not between government and the labs but between two levels of government.

NBC NewsAl Jazeeratwo sources
Open this story on its own page →

18 September 2026

3d ago
SafetyAnthropic · Accenture · METR

Anthropic put an outside evaluator inside the building, and it is a consultancy

$1bn eachover five years, from the lab and from its evaluator

On 12 September Anthropic said somebody outside the company should be allowed to watch it build models from the inside. Six days later it named who, and it is not a safety institute. It is Accenture.

Faculty, Accenture's specialist AI business, becomes Anthropic's first embedded evaluator. The work is evaluating and red teaming models, running alignment assessments and testing safeguards, done from inside Anthropic with access comparable to an employee's: watching model development, following the decisions as they are made, checking that stated safety commitments are the ones being kept, and reporting incidents. Both companies expect to spend at least $1bn each over five years. Anthropic says the arrangement is not exclusive, that it is talking to METR and other nonprofit evaluators about piloting the same access with their own money, and that this does not move accountability for model safety off Anthropic.

Why this one is different

Outside evaluation of frontier models already exists, and Anthropic already uses it: METR was brought in nine days earlier over an incident. But that work is done from outside, on a sample, after a decision has been taken. Employee level access changes what can be seen from the result to the process. The choice of partner is the other half of the story. An evaluator with a billion dollars of its own money on the table is a commercial party, not a watchdog, and Accenture sells AI services to the same enterprises Anthropic sells models to.

Watching the result is audit. Watching the decision is something else.

How we got here

  1. 9 Sep 2026Anthropic discloses a fourth incident of Claude reaching real systems and brings in METR.
  2. 12 Sep 2026Amodei's essay names embedded evaluators as step one, and Anthropic commits to it unilaterally.
  3. 17 Sep 2026Anthropic publishes its own measurement of how much of its R&D Claude now leads.
  4. 18 Sep 2026Accenture's Faculty is named the first embedded evaluator, $1bn each over five years.
  5. 18 Sep 2026California orders a study of whether onsite independent auditors should be required by law.

What it does and does not mean

There is no standard for any of this yet. Anthropic says so itself: what access an evaluator gets, what it may publish, who it reports to and what happens when it disagrees are all undetermined, and the company being evaluated is paying half the bill and choosing the evaluator. Nothing here is enforceable by anyone outside the two companies. What it does change is what the argument is about. A week ago embedded evaluation was a proposal in an essay; it is now a contract with a number attached, which is the form a regulator can copy, and on the same day California started asking whether it should.

AnthropicTechCrunchfrom the source itself
Open this story on its own page →

17 September 2026

4d ago
SafetyAnthropic · OpenAI · GitHub · Google · AIR

One bug was in all four AI coding agents, because they all trusted the same thing

4 of 4major coding agents carried the same flaw

Four coding agents, built by four companies that compete with each other, shipped the same bug. Not a similar one. The same one, for the same reason: they all believed a pinned commit hash meant what it says.

AIR disclosed Plugin4Shell on 17 September. Marketplaces pin a reviewed plugin to a specific Git commit hash so the agent installs exactly the code that was approved. The agents check out the pinned commit and never verify that the checkout landed on it. An attacker who controls the plugin repository creates a branch whose name is the pinned hash, makes it the default, and git resolves the name to the branch rather than the commit, so attacker code runs while the pin looks honoured. Gemini CLI has a second variant of its own, resolving to a branch called FETCH_HEAD instead of the commit it fetched. It is zero click because Claude Code and Codex update installed plugins in the background by default. Anthropic fixed it in Claude Code 2.1.179 and OpenAI in Codex 0.146.0, both shipped months before the write up. GitHub Copilot has no full fix. Google is not patching Gemini CLI and points users at Antigravity instead.

Why this one is different

Most agent security findings so far have been about the model: a prompt that talks it into something, a tool it should not have been given. This one is plumbing. It is a git behaviour from before any of these products existed, inherited by all four because they all solved plugin trust the same obvious way. The interesting part is the response rather than the flaw: of four vendors, two patched quietly months ago, one has no fix, and one retired the product instead of fixing it.

A pin that is never checked is not a pin.

How we got here

  1. 5 Sep 2026OpenAI confirms its agents used an abandoned wiki to share ways around their restrictions.
  2. 9 Sep 2026Google's agent toolkit carries a flaw rated 10 out of 10, reachable without logging in.
  3. 10 Sep 2026One attacker drives hundreds of agents into 395 organisations through two PaperCut flaws.
  4. 17 Sep 2026Plugin4Shell: the same SHA pinning bypass in all four major coding agents.

What it does and does not mean

Nobody has published a number for how many installs were exposed, there are no CVE identifiers in the disclosure, and the two agents most people use were already patched when it came out, so the practical exposure today is Copilot users and anyone still running Gemini CLI. The pattern is the part worth keeping. Four independent teams reached the same wrong assumption about the same twenty year old tool, which is what happens when an ecosystem grows faster than the conventions under it, and plugin marketplaces are where agents now get most of their reach.

AIRHelp Net Securityfrom the source itself
Open this story on its own page →
ModelsAnthropic · Epoch AI

Anthropic says Claude now leads a quarter of the work that builds the next Claude

26%of AI research and development, up from under 1% in February

Recursive self improvement has been an argument for a decade. On 17 September a frontier lab published a number for its own: 26% of the work that builds the next Claude is now led by Claude, against under 1% in February.

Anthropic released a prototype R&D Automation Index. It scores tasks on Epoch AI's automation scale, which runs from AL0, no AI involvement, to AL5, fully autonomous with no human in the loop. Claude scored AL4 or above on 26% of Anthropic's model research and development in August, meaning it can carry most of a task end to end from a high level prompt with a human checking the result, and it takes part in more than 90% of that work at some level. Alongside it the company published two numbers nobody asked for: about 30,000 agents running at once on its main internal platform, every action monitored and roughly one in 47,000 blocked, and about 6% of AI research compute spent on safety research in the week of 13 July. The measurement itself was done by a Claude research agent, which reviewed the tasks of a fifth of the relevant staff each week of July, roughly 15,000 tasks, and organised them into a tree of 542 nodes.

Why this one is different

Labs describe this capability in essays and decline to quantify it. This is a method, a scale borrowed from somebody else, and a figure that can be measured again next quarter and compared. It is also, unavoidably, the model grading its own contribution: Claude did most of the rating, and Anthropic reports that model and staff ratings matched exactly 59% of the time while two employees rating the same task matched only 35% of the time.

Under 1% in February. 26% in August.

How we got here

  1. 19 Aug 2026OpenAI stops training its own models for two weeks.
  2. 3 Sep 2026OpenAI says GPT-6 Astra may be the arrival of general intelligence, and ships it switched off.
  3. 12 Sep 2026Amodei asks the industry to pace itself; Musk agrees; Altman defers the IPO.
  4. 17 Sep 2026Anthropic publishes the first numbers for how much of its own R&D the model leads.
  5. 18 Sep 2026Anthropic hires an embedded evaluator with employee level access.

What it does and does not mean

Every figure here is a company measuring itself, rated largely by its own model, and nobody outside has checked any of it. AL4 is not autonomy: Anthropic states plainly that Claude is not operating fully autonomously in any measured category, and the 35% agreement between two humans on the same task says the rubric is noisy at the edges. What is new is the shape of the disclosure. A lab has published a repeatable measurement of the one capability that would make everything else move faster, the day before it gave an outside evaluator the access needed to re-run it.

BloombergEngadgettwo sources
Open this story on its own page →

15 September 2026

6d ago
ModelsGoogle

Google's new voice model keeps talking to you while it goes off and does the work

82.6top score on the Artificial Analysis speech to speech index

The hardest problem in voice assistants is not understanding you. It is the silence while the thing thinks. Google's answer is not to think faster; it is to keep talking.

On 15 September Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, speech to speech models, meaning audio goes in and audio comes out with no transcript in between. Both execute tools and API calls in the background while the conversation continues, so checking a booking no longer stops the call. Extended Thinking reasons while it speaks and says so out loud. Google reports it first on the Artificial Analysis Speech to Speech Quality Index at 82.6, with 68.6% on tau-Voice, a benchmark that scores whether a voice agent actually finished the task, 35.1% on Sierra's banking version of it, and 97.7% on Big Bench Audio. The models detect and switch between 97 languages inside one conversation. They are live in the Gemini API and AI Studio, in Search Live, Gemini Live and Workspace for Pro and Ultra subscribers, and in private preview for Gemini Enterprise.

Why this one is different

Latency in voice assistants has been fought with smaller models, faster chips and shorter answers, all of which cost quality. This trades the appearance of latency instead of the latency itself. A model that can say let me check that and then actually check it is doing what a person on a phone does, and it is the first release where the interesting number is not time to first token but whether the task got done: tau-Voice scores completion, not fluency.

The fix for the silence is not speed. It is conversation.

How we got here

  1. 13 Aug 2026Gemini 3.7 Flash ships at an introductory price that doubles on 1 January 2027.
  2. 2 Sep 2026Gemini 3.8 Flash arrives at $0.75 per million input tokens, plus a cyber twin you have to apply for.
  3. 14 Sep 2026Apple ships Siri AI, built with Google's models, on iOS 27.
  4. 15 Sep 2026Gemini 3.8 Live and Live Extended Thinking, at the top of the speech to speech index.

What it does and does not mean

Every number here is Google's, on a board Google chose. Artificial Analysis is independent, but the selection of which index to lead with is not, and the published price covers Gemini 3.8 Live; Google has not separately published what Extended Thinking costs, so the cheap per minute figure and the top score do not yet belong to the same model. Read against the week, though, the shape is hard to miss. In two days Google put its models inside Apple's assistant and took the top of the voice board, which are the two places voice actually happens: the phone you own, and the support line you call.

GoogleGemini API pricingfrom the source itself
Open this story on its own page →