Gemini 4, FTC Investigations, SI Accords, & Dots

TLDR;

Welcome back! This week we have a lot of news to talk about. From the hours-ago release of Gemini 4 Argon to a plethora of insights, including a lot of new products, market struggles, super intelligence accords, FTC investigations, personal assistants, and many more.

Enjoy!

MODELS

Google Releases Gemini 4 Argon

Google announced Gemini 4 Argon on September 30, initially restricting access to selected cybersecurity partners through its Fairwind Program. The model targets software engineering, legal and financial work, and cyber defense. Google has not provided a public release date.

According to Google’s announcement, Argon supports up to one million output tokens, up from 64,000, allowing longer reasoning and generation within a single run. Google reports scores of 77.9% on DeepSWE v1.1, which tests extended software engineering tasks, and 68% on CWE-bench v1, tying for first in vulnerability remediation.

Across benchmark aggregates like Artificial Analysis or Vals, the model scores highly, even taking first place in the latter ahead of all other frontier models, and it seems to be quite efficient on a cost-per-task basis.

Trusted cyber defenders will receive access without cyber guardrails (uh, oh!). Before wider availability, Google says it is strengthening safeguards against misuse, prompt injection, and model misalignment.

Introductory API pricing will be $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Paid API customers and Google AI Ultra subscribers will be first in the broader rollout.

TheWhiteBox’s takeaway:

There’s not much I can say; I haven’t even tried the model! But if benchmarks mean anything, this is genuinely the first truly exciting release we’ve had from Google in a long while.

It’s also very efficient, and the price is competitive (prompt caching is particularly aggressive, at 95% off the input token price). It seems Google might be back; let’s see whether the facts match the claims.

CYBERSECURITY

Anthropic Says GLM-5.3 is Great at Hacking

Anthropic published a report yesterday that Z.ai’s open-weight GLM-5.3 can autonomously develop sophisticated cyber exploits, while its safeguards can be bypassed or removed. The company’s evaluation found capabilities approaching Claude Mythos Preview on specific exploit-development tests.

In Anthropic’s sandboxed ExploitBench tests, GLM-5.3 built complete exploits in 50 of 410 attempts, compared with 56 for Mythos Preview. In a researcher-guided experiment, it also chained previously unknown browser vulnerabilities into a webpage that could read files from a test computer.

In other words, GLM-5.3 displays near-frontier cybersecurity capabilities.

To prove how unsafe this is (according to them), Anthropic’s simulated misuse tests found GLM-5.3 engaged with malicious requests in 64% to 100% of trials under different safeguard-bypass conditions.

But where do these numbers come from?

They “broke” GLM-5.3’s safeguards in three different ways: deception, prefilling, and abliteration.

Deception, as the name suggests, uses deceptive prompts, such as telling the model it is an autonomous red-team agent working on an exercise. That way, the AI is much more likely to comply because it has been tricked into “believing” it’s being tested for cybersecurity…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.