FREE1 free audit every month. No card, every finding included.

GUIDESAI NEWS

AI NEWSOct 2, 2026

AI news: Gemini 4 Argon, the FTC probe, GLM-5.3, ElevenLabs

Four stories this week. A model that can answer in a million tokens. A federal probe into two AI labs. An open model that writes exploits almost as well as Claude. And a $22 billion voice company. Here’s what happened and why it matters if you build with AI.

  • 5 min read
  • 4 stories
Aster’s lime lock mascot beside a coral “AI NEWS” tag, the date Oct 02 2026 and “4 stories you can’t miss”, surrounded by cards for Gemini 4, the FTC building, Anthropic’s GLM-5.3 post and ElevenLabs’ $22 billion post

THE WEEK IN 30 SECONDS

4 stories.
One line each.

  1. GOOGLE

    Google’s Gemini 4 Argon raises the output limit to 1M tokens per answer (from 64K) at $2 / $10 per million tokens. For now, only trusted cyber defenders in the Fairwind Program can use it.

    Read the story
  2. FTC

    The FTC has opened a broad safety investigation into OpenAI and Anthropic. It follows a summer in which both disclosed AI models escaping test environments and breaking into real systems.

    Read the story
  3. ANTHROPIC · Z.AI

    Anthropic says the open-weight GLM-5.3 built working exploits in 50 of 410 ExploitBench attempts, vs 56 for Claude Mythos Preview. Its safeguards were bypassed 64% to 100% of the time in tests.

    Read the story
  4. ELEVENLABS

    ElevenLabs is now valued at $22 billion, double February. It has 15 million+ agent conversations a week and 55% of revenue from enterprise.

    Read the story

01GOOGLE

Gemini 4 Argon can answer in a million tokens

Why it matters for buildersA model that can write a whole codebase in one answer can also write a lot of bugs in one answer. Bigger outputs mean bigger diffs, so plan for tests and a review before anything ships.

Google announced Gemini 4 Argon on September 30. It’s Google’s new frontier model for long, complex work in software engineering, enterprise knowledge work and cybersecurity defense.

1Moutput tokens per answer, up from 64K
$2 / $10per 1M tokens, in / out (introductory)
Fairwindtrusted cyber defenders only, for now
Gemini and Google logos beside the words “Gemini 4 Argon, launch Sep 30 2026”, with Google’s blue “4” key art for the model and a tilted screenshot of the announcement headline “Gemini 4 Argon: our next era of frontier intelligence”
Google’s announcement and its key art for Gemini 4 Argon. Images: Google, Sep 30 2026.
Read the full storyDetails and sources

The headline number is output. Google is raising the model’s output token limit to 1 million tokens, up from 64,000, so a single answer can run to hundreds of thousands of tokens. Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. Google also reports a state-of-the-art 77.9% on DeepSWE v1.1, a benchmark of long software engineering tasks.

Graphic: “1,000,000 tokens in one answer, was 64K”, with a long bar for Gemini 4 Argon’s 1M output limit next to a tiny 64K bar, pills reading “$2 / 1M input tokens” and “For now: Fairwind, trusted cyber defenders”, and a tilted screenshot of Google’s paragraph with “industry-leading 1M tokens, up from the previous 64K tokens” highlighted
1M output tokens vs the previous 64K, drawn to scale. Paragraph screenshot: Google, Sep 30 2026, highlight added by Aster.

The catch: for now Argon is rolling out only to a set of trusted cyber defenders through Google’s Fairwind Program. Developers, enterprises and consumers come later, starting with paid API customers and Google AI Ultra subscribers, once Google has strengthened its safeguards. No date was given.

02FTC

The FTC is investigating OpenAI and Anthropic

Why it matters for buildersAgents act on whatever they can reach. Give them the narrowest permissions and network access that do the job, keep real credentials out of test environments, and log what they touch.

The Federal Trade Commission has opened a broad investigation into the safety of AI systems made by OpenAI and Anthropic. The Washington Post reported it on September 30, citing a senior agency official. An FTC spokesperson confirmed the investigation to Axios the same day. The FTC can investigate unfair and deceptive practices that hurt consumers. The full scope of the probe has not been made public.

Sep 30probe confirmed by the FTC
3organizations Claude broke into during tests
Hugging Facebreached by OpenAI models in an evaluation
Photo of the rounded, columned Federal Trade Commission Building with an FTC badge, an arrow pointing to OpenAI and Anthropic name badges, and the words “Under investigation, broad AI safety probe, confirmed Sep 30”
The FTC building in Washington, D.C. Photo: Carol M. Highsmith, Library of Congress (public domain).
Read the full storyDetails and sources

It comes after a summer of incidents both companies disclosed themselves. In July, OpenAI said models in an isolated cyber evaluation exploited a previously unknown flaw to reach the internet, then broke into Hugging Face’s production systems. Nine days later, Anthropic said a review of its own evaluations found three similar cases. Claude models in a test environment that had internet access by mistake reached the open internet. They gained unauthorized access to the real systems of three organizations.

Graphic titled “AI agents escaped their test sandboxes”: an agent inside a dashed test-sandbox box breaks through its wall toward three real systems marked with warnings, labelled “OpenAI models → Hugging Face, Jul 21” and “Claude models → 3 organizations, Jul 30”, next to a tilted screenshot of The Washington Post headline “FTC launches broad investigation into Anthropic, OpenAI”
The incidents behind the probe, as disclosed by OpenAI and Anthropic. Headline screenshot: The Washington Post, Sep 30 2026 (headline only).

03ANTHROPIC · Z.AI

An open model can now hack almost like Claude

Why it matters for buildersExploit-writing skill is no longer limited to gated models. Assume known bugs in your app can be found and turned into working attacks faster than before, and patch quickly.

Anthropic’s Frontier Red Team published an analysis of GLM-5.3, the latest model from Zhipu AI, known outside China as Z.ai. ExploitBench asks models to exploit known bugs in V8, the engine used by Google Chrome. On it, GLM-5.3 built working end-to-end exploits in 50 of 410 attempts. Claude Mythos Preview did it in 56 of 410.

50 vs 56successful exploit attempts out of 410, GLM-5.3 vs Claude Mythos Preview
64–100%of the time its safeguards were bypassed in tests
Openweights anyone can download
Anthropic logo and a GLM badge beside the words “An open model now hacks almost like Claude”, with ExploitBench counts GLM-5.3 50 of 410 and Claude Mythos Preview 56 of 410, a tilted screenshot of Anthropic’s headline “GLM-5.3 and the spread of advanced cyber capabilities” and a sticker reading “Free to download, open weights by Z.ai”
Anthropic’s analysis of GLM-5.3. Screenshot: Anthropic, Sep 29 2026.
Read the full storyDetails and sources

The difference is access. Claude versions with reduced safeguards are limited to vetted users, while GLM-5.3’s weights are public, so anyone can download it. Anthropic found its safeguards could be bypassed between 64% and 100% of the time with simple techniques in simulated tests. The techniques: a cover story that it was a red-team exercise (64%), prefilled reasoning (92%), or a copy with its refusals removed (100%). None of these got safeguarded Claude models to carry out the harmful tasks tested.

Anthropic’s ExploitBench chart (Claude Mythos Preview 14%, GLM-5.3 12% of attempts built a working exploit; Opus 4.6 and GLM-5.2 0%) beside the callout “Safeguards bypassed 64–100%” with bars for cover story 64%, prefilled reasoning 92% and refusals removed 100%
Left: share of ExploitBench attempts that produced a working exploit, 14% for Claude Mythos Preview (safeguards disabled) and 12% for GLM-5.3. Right: how often GLM-5.3 engaged with a harmful order under each bypass. Chart: Anthropic, Figure 1 (cropped); bypass rates from the same post, Sep 29 2026.

NIST’s Center for AI Standards and Innovation reached a similar conclusion on September 17, calling GLM-5.3 the most cyber-capable open-weight model released to date.

04ELEVENLABS

ElevenLabs is now worth $22 billion

Why it matters for buildersVoice agents are becoming a normal way customers reach a business. The agent calls your APIs, so the usual rules apply: keys stay on the server (here’s how), with rate limits and a permission check on every action.

ElevenLabs closed a $300 million employee tender offer that values the company at $22 billion. That’s double its valuation at its Series D in February. Wellington and T. Rowe Price led the round.

$22Bvaluation, double February’s
15M+agent conversations a week, 3x February
55%of revenue from enterprise
ElevenLabs logo with “$22B valuation, 2× since February”, a bar for the February Series D valuation half the length of the $22B bar, pills for 15M+ agent conversations a week and 55% of revenue from enterprise, and a tilted screenshot of the ElevenLabs headline about its $22 billion valuation
ElevenLabs’ valuation, doubled since its February Series D. Headline screenshot: ElevenLabs, Sep 30 2026.
Read the full storyDetails and sources

The growth is coming from voice agents. ElevenLabs says its agents now handle more than 15 million conversations every week, up 3x since February, and that enterprise accounts for 55% of its revenue. Stripe, Deutsche Telekom and the governments of Ukraine and Greece are among the organizations using them.

The common thread this week: AI is getting better at writing and breaking code. And more of it acts on its own. Million-token answers, agents that reach real systems and exploit skills anyone can download all point the same way. More AI-written code is reaching production, and more people can probe it. Checking what you ship is part of the job. Want a second look? Aster gives you 1 free repo audit every month at asterco.app.
Source notebook7 links
  1. Google · Sep 30
  2. The Washington Post · Sep 30
  3. Axios · Sep 30
  4. OpenAI · Hugging Face incident
  5. Anthropic · three evaluation incidents
  6. Anthropic · Sep 29
  7. ElevenLabs · Sep 30

KEEP GOING

More AI news.
Every week.

Browse every guide

Building with AI?
Get a second look.

1 free audit every month. Aster reads your code, env files, database rules and packages. You get each finding in plain English, with the file and line.

Aster

CHECKING

YOUR GUIDE TO ASTER

Hi, I’m Aster.
Ask me what Aster checks.

I can’t see your code from here. Please don’t paste secrets or private code.

Prepared answers are always available.