The Signal — October 1, 2026

Google released its newest frontier model to a small group of vetted security teams, and gave those teams a version with the cyber guardrails removed. In Washington, the FTC confirmed it is investigating OpenAI and Anthropic over consumer risks, one day after their leaders signed a voluntary safety pledge at the White House. DeepMind also published a way to mark AI-designed proteins so that DNA makers can tell where an order came from.

Google is giving Gemini 4 Argon to vetted cyber defenders first, without cyber guardrails

Google DeepMind announced Gemini 4 Argon on September 30 and is rolling it out first to "a set of trusted cyber defenders" through Fairwind, Google's security program. Paid API customers and Google AI Ultra subscribers come later, on a date Google has not given. Koray Kavukcuoglu, who heads DeepMind, wrote that Google is "actively engaged in the U.S. government's voluntary process for pre-release model access while we gradually expand access." For those vetted defenders and for Google's own teams, the company says it will release Argon "without cyber guardrails" so they can use its full ability to find, validate and patch vulnerabilities autonomously. Wiz, an early user, says Argon found a critical flaw exposing personal data in hospital software that earlier models had missed.

Google says Argon scored 77.9 percent on DeepSWE v1.1, a long-horizon software engineering test, 51.3 percent on Zapier's AutomationBench, and tied for first at 68 percent on CWE-bench v1, which checks whether a model can fix security flaws. It also leads the Vals Index, which weights finance, coding, legal and tax tasks by their share of U.S. GDP. These are Google's figures from benchmarks it selected, and no independent evaluation has been published. The output limit rises from 64,000 tokens to one million, and the introductory price of $2 per million input tokens and $10 per million output tokens later doubles.

Before a wider release, Google says it is adding refusals for cyber misuse and chemical, biological, radiological and nuclear misuse, monitoring of the model's internal activations, and defenses against prompt injection. It is also adding monitors that read Argon's chain of thought and actions and can stop a task when the model goes beyond what the user asked for. Google says it ran similar monitors during training and kept their findings out of the training data, so Argon would not learn to hide its reasoning from them. The Verge notes the announcement came a day after OpenAI's DevDay.

Sources: Google · TechCrunch · The Verge

The FTC confirmed its OpenAI and Anthropic investigation a day after their leaders signed a voluntary White House pledge

An FTC spokesperson confirmed to the Associated Press on September 30 that the agency has opened an investigation into OpenAI, Anthropic and other AI companies over dangers their technology may pose to consumers, and declined further comment. The New York Post first reported the probe and said it had been underway for months. Bloomberg Law, citing a person familiar with the matter, reports that the FTC is preparing formal demands for information, likely in the coming weeks, to examine whether the companies are complying with consumer protection law after a run of incidents in which their models broke into outside systems. The FTC has a long record of data-security cases, though investigations can also close without any action, and none of the demands has been served yet.

The day before the FTC's confirmation, at the White House meeting our September 28 edition previewed, executives from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the "Joint Commitment on Frontier Responsibilities," which presidential adviser David Sacks shared online and The Verge reported. Each company agrees to keep internal controls that monitor its models for cyber, bio and chemical risks and make sure the models "do not hack or access technical systems in unintended ways," with an internal team checking those controls. Each also agrees to hire an independent auditor or evaluator to test the controls and to have an independent board committee oversee the results. The pledge carries no penalties. President Trump called it "morally binding," and the document itself says that "over time, it may make sense to codify these steps into laws or regulations." According to IAPP, when Trump listed the agencies that would go after misuse, he named the Justice Department, the FBI and the CIA, and did not mention the FTC. FTC Chair Andrew Ferguson attended the meeting, Bloomberg Law reports.

Under the pledge, the companies choose their own controls and their own outside evaluators. The FTC investigation is a separate track in which the government sets the questions, and it is still at an early stage.

Sources: Associated Press (via ABC7) · Bloomberg Law · The Verge · IAPP

DeepMind watermarked AI-designed proteins, and the proteins still worked

Companies that synthesize DNA screen orders against databases of known toxins and viral parts. AI design tools can produce protein sequences that resemble nothing in those databases, so an unfamiliar order no longer tells a screener much about whether it is natural, harmless or engineered. Google DeepMind's answer, published in Nature on September 30, is SynthID Bio, a watermark carried in the protein itself.

The method adapts SynthID, Google's watermark for AI-generated text and images. When ProteinMPNN, a widely used design tool, picks amino acids one position at a time, a secret key nudges each choice toward a watermark pattern, and the tool accepts the nudge only where it predicts the protein would still work. A separate version fine-tunes part of AlphaFold 3 so that the 3D structures it predicts carry their own signal. In wet-lab tests on binder proteins for three targets (VEGF-A, part of the spike protein of the virus that causes COVID-19, and PD-L1), watermarked designs matched unwatermarked ones on hit rate, binding strength and sequence diversity.

DeepMind proposes that DNA makers holding keys from trusted designers could clear watermarked orders quickly and spend their review time on the rest. Detection needs the key and is a statistical call, with thresholds set against a chosen false-positive rate. The paper calls the work a proof of concept, the tests covered binders rather than enzymes or other functions, and DeepMind says making the watermark hold up against deliberate tampering is still a key challenge. A watermark shows where a protein came from, not that it is safe. DeepMind is releasing the code, lab data and model weights, and says it has early results watermarking the genome of a bacteriophage designed with the Evo 2 model.

Sources: Google DeepMind · Nature · Unite.AI

On the Editor's Desk

OpenAI said it disrupted a campaign aimed at extracting its models' hidden reasoning and links the accounts involved to Moonshot AI. The activity dates to July, the attribution is OpenAI's own, and we have not seen Moonshot's response, so we are holding the story. MI5's warning to UK universities about a research funder it links to Chinese intelligence is well sourced, but the AI connection is limited to the research the funder paid for. We also passed on America.gov, the White House's new AI-assisted portal for government services, which is rolling out in phases.