Sakhanda Wire
NVDA $153.62 -1.62% MSFT $506.90 -2.14% GOOGL $279.81 -1.38% META $619.78 -1.06% AMZN $256.58 -0.45%

The Decoder

625 stories · all sources

The Decoder

Open and closed AI models show narrowing gap in cyber capabilities

The Vals CyberOffense benchmark found that AI agents can independently attack simulated corporate networks, starting with only a public web address. Running from isolated Kali Linux machines, the agents move through web applications and Windows servers toward administrator access, while an external system records and scores every action. GPT-6 Sol Daybreak achieved the strongest result, reaching 80.8 percent of the objectives. The open-weight Deepseek V4.1 Flash 1017 reached 72.8 percent, at about $3 per run compared with roughly $22 for GPT-6 Sol Daybreak.

The Decoder

Fine-tuning cuts propaganda in Alibaba’s Qwen3.8-27B to 0.9 percent

Nous Research says it retrained Alibaba’s open-weight Qwen3.8-27B to reduce responses reflecting Chinese state propaganda. The resulting model, Arete-27B, was trained with supervised fine-tuning on about 53 million tokens over two epochs. The process used neither reinforcement learning nor preference optimization, and the team removed overlaps between training and test prompts.

The Decoder

Google's AI screens for high blood pressure using ordinary smartwatches

Google Research has developed an artificial intelligence system that looks for signs of high blood pressure using only the pulse and motion sensors found in modern smartwatches. The approach does not require a blood-pressure cuff for calibration. The study was published in Nature.

The Decoder

OpenAI secures $11 billion in first tranche of funding round

OpenAI has received about $11 billion in new capital, according to Bloomberg, marking the first tranche of a financing round that values the company at $1.25 trillion. The funds were transferred on Friday, while the overall round could reach $20 billion, with additional money expected in the coming weeks.

The Decoder

Ripple-65B Preview posts strong coding benchmark results

Ricursive Intelligence has released Ripple-65B Preview, an open-weight language model with 65 billion parameters. The AI chip design company says it was pretrained on 12 trillion tokens and reached 80.52 on TerminalBench 2.1, which it describes as the highest score from an open U.S. model. The stated total cost for research, development and training was $32 million.

The Decoder

Perplexity introduces always-on AI agents for ongoing projects

Perplexity has introduced Teammates, AI agents designed to remain active around the clock and handle continuing work independently. The new feature differs from Perplexity Computer, which carries out individual assignments. Each Teammate operates inside its own virtual machine equipped with a desktop and browser, allowing it to work through a project without constant user supervision.

The Decoder

Thinking Machines unveils Inkling 2 Small open-weight model

Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has introduced Inkling 2 Small, a 594-billion-parameter mixture-of-experts model. About 17 billion parameters are active at a time. The model accepts text and images and allows users to adjust its thinking depth. Thinking Machines describes it as the strongest Western open-weight model, although it is initially available as a proprietary service through the Tinker platform.

The Decoder

Anthropic faces investor pushback over planned $1.5 trillion IPO

Anthropic’s planned initial public offering is facing resistance from potential investors, who question whether the company can justify a $1.5 trillion valuation. Reuters reports that backers are uncertain whether the scheduled roadshow can begin next week. In September, Anthropic was considering a valuation of as much as $2 trillion, while one investor has reportedly suggested $1 trillion instead.

The Decoder

Claude adds read-only access to users’ bank and card accounts

Anthropic is introducing Personal Finance, a Claude feature that lets users connect bank and card accounts and ask questions about their finances. Claude receives read-only access, so it cannot move, manage, or invest money. The connection is handled through Plaid, meaning Claude does not see the user’s bank password.

The Decoder

OpenAI introduces GPT-6 Cyber for vulnerability research

OpenAI has unveiled GPT-6 Cyber, a model trained to identify software vulnerabilities, create working exploits and propose patches. The company says it found 81 previously unknown severe vulnerabilities in Firefox; Mozilla accepted the reports and has released fixes. The model also reassessed 30 known Firefox issues as more serious than previously believed.

The Decoder

Google launches Gemini Omni 1.2 Flash for 15-second video clips

Google DeepMind has made its Gemini Omni 1.2 Flash video-generation model generally available. The model produces clips from three to 15 seconds, extending the previous 10-second limit, and costs $0.07 per second for 720p video, 30 percent less than its predecessor. It is being offered through the Gemini API, Google AI Studio, Google Cloud Console, Flow, Google Vids, Google Ads, and the Gemini app for Google AI Ultra subscribers. Each generated video includes an imperceptible SynthID watermark.

The Decoder

Baidu releases 280-billion-parameter ERNIE-5.5-Flash-Base

Baidu has released ERNIE-5.5-Flash-Base as an open foundation model with 280 billion parameters. The mixture-of-experts system activates 14 billion parameters for each input. It was pretrained on 33.5 trillion tokens with a training context of up to 128,000 tokens.