OpenSight News: AI

OpenAI releases GPT-6 Cyber and rates its hacking skill as critical

Ari — Sat, 07 Nov 2026 00:00:00 GMT

OpenAI has released GPT-6 Cyber, a model built to find software flaws and write working exploits. It is the first model OpenAI rates as critical for cybersecurity under its own safety framework. Because of that rating, only vetted organisations in its Daybreak Red programme can use it. They must verify identity and meet strict account security rules.

On the ExploitGym benchmark, which tests turning known bugs into working attacks, the model succeeded in about 48 percent of tasks. Its predecessor managed about 32 percent. OpenAI also rates it high for biological and chemical risk. Safeguards include sandboxed use, monitoring for misuse, and blocking help against industrial control systems and essential services. The benchmark figures come from OpenAI itself and have not been checked independently.

Sources

GPT-6 Cyber System Card — OpenAI, 5 Nov 2026 (primary source)

METR finds Anthropic is improving its models about twice as fast as before

Ari — Sat, 07 Nov 2026 00:00:00 GMT

The independent evaluator METR says Anthropic is improving its AI models about twice as fast as it did between early 2024 and late 2025. METR estimates that most of the speed-up comes from AI tools helping Anthropic's own staff with coding and experiments. More training compute is not the main driver, it says. METR found no clear bottleneck that would stop further acceleration within a year.

The assessment covered August to October 2026. Its central estimate is a speed-up of about one point nine times, with a wide range of uncertainty. Anthropic withheld most of the evidence METR found most useful, saying it could help competitors. METR also notes that its benchmark measure may not match real-world impact, and that the earlier rate may have been unusually slow. Staff reported large personal productivity gains, but those figures are self-reported.

Sources

Assessment of AI R&D acceleration at Anthropic — METR, 5 Nov 2026 (primary source)

OpenAI says its AI agents built working root exploits for 948 Linux kernel bugs

Ari — Fri, 06 Nov 2026 00:00:00 GMT

OpenAI says an unreleased AI model for cyber security built working root exploits for 948 different Linux kernel bugs. It did so in a single automated run in October that cost a little over one million dollars in compute. That is roughly 1,300 dollars per exploit. The company says the agents also found at least ten new attack techniques.

The pipeline first flagged about 2,300 possible bugs and confirmed about 1,700 of them. Security staff review every exploit and proposed patch before it goes to the kernel maintainers. OpenAI says hundreds of reports and fixes have gone out so far, starting with the most serious, and partners in its Patch the Planet programme are sending the rest. Its warning for operators is blunt. On an unpatched Linux system, any low-privilege access should be treated as close to full control for a capable attacker. The work shows how cheap exploit finding has become, for defenders and attackers alike.

Sources

A thousand fewer paths to root — OpenAI, 5 Nov 2026 (primary source)
OpenAI's upcoming AI model built 948 working Linux root exploits in a single run — Help Net Security, 5 Nov 2026
AI Agents Got Root on Linux via 948 Kernel Flaws in OpenAI Research — SecurityWeek, 5 Nov 2026

Microsoft passes on an option to more than double its Texas AI capacity deal with Nscale

Ari — Fri, 06 Nov 2026 00:00:00 GMT

Microsoft will not take up an option to more than double its rented AI computing capacity at an Nscale data centre in Texas. That is according to a person familiar with the matter, first reported by the Financial Times. The option covered a second phase of about 700 megawatts from late 2027. It would have been worth tens of billions of dollars to Nscale.

The existing Texas contract of about 230 megawatts of Nvidia systems is not affected. Nscale has put the total value of its Microsoft contracts at about 103 billion dollars, and it is in talks with other large customers about the rest of the site. The decision was reportedly made months ago. It follows Microsoft's exit from Nscale's planned UK site in Essex, where grid connection delays pushed supply into the 2030s. Neither company commented. The news landed as investors were already worried about a slower data-centre buildout.

Sources

Microsoft declines option to expand Texas data center deal with Nscale, source says — Reuters, 4 Nov 2026 (primary source)
Microsoft declined option for expanded compute contract in Texas with Nscale — DatacenterDynamics, 5 Nov 2026

Marvell agrees to buy Lumentum for about 115 billion dollars to own AI optics

Ari — Tue, 03 Nov 2026 00:00:00 GMT

Chip designer Marvell has agreed to buy Lumentum, a maker of lasers and optical parts, in an all-stock deal. The deal values Lumentum at about 115 billion dollars including debt. Lumentum holders get four and a half Marvell shares for each share they own. The two companies say AI data centres need far more bandwidth than copper cables can carry, so links are moving to light.

The price is about 24 percent above Lumentum's share price before talks were reported. Marvell holders would own about two thirds of the combined company. The firms expect more than one billion dollars a year in cost savings. Closing needs approval from both sets of shareholders and from regulators, and is planned for the first half of 2028. Marvell shares had fallen sharply when the talks first leaked last week.

Sources

Marvell and Lumentum to Combine, Creating a Category Leader in AI Connectivity — US Securities and Exchange Commission, 2 Nov 2026 (primary source)
Stock market news for Nov. 2, 2026 — CNBC, 2 Nov 2026

Google launches Gemini 4 Carbon for security work but only cyber defenders get it

Ari — Mon, 02 Nov 2026 00:00:00 GMT

Google has released Gemini 4 Carbon, a model built for software security. It can look for vulnerabilities, test whether they can be exploited and write a fix. Google calls it its strongest model so far. For now only selected cyber defenders can use it. Governments, critical infrastructure operators and large platforms come first, through a programme called Fairwind.

The limit has a reason. A model that finds and exploits flaws for a defender can do the same for an attacker. Google reports 88 percent on a benchmark for finding vulnerabilities and 71 percent on one for patching them. These scores are Google's own. The introductory price is two dollars per million input tokens and ten dollars per million output tokens, doubling later. No date for general access has been given.

Sources

Introducing Gemini 4 Carbon — Google, 28 Oct 2026 (primary source)
Gemini 4 Carbon cyber defenders — BetaNews, 28 Oct 2026

Mistral publishes the weights of its trillion-parameter Large 4 model

Ari — Sun, 01 Nov 2026 00:00:00 GMT

Mistral, the French AI company, has released the full weights of Mistral Large 4. The model has about one trillion parameters. Anyone can now download it and run it on their own hardware. Mistral says it is the strongest open model for coding and for security work such as finding and fixing vulnerabilities.

The weights are on Hugging Face under a modified MIT licence. Use, fine-tuning and commercial use are allowed. Companies that want to sell the model as a hosted cloud service need a separate licence. The release follows a preview earlier in October, and Mistral says three more weeks of training lifted its terminal benchmark score from 28 to 41 percent. The benchmark claims are the vendor's own. For teams that need to keep data in house, it widens the choice of large models they can host themselves.

Sources

Mistral Large 4 (Le Chonk) is now open weight — Mistral AI, 29 Oct 2026 (primary source)
Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity — TechCrunch, 29 Oct 2026

Apple Intelligence comes to China in December, powered by Alibaba's Qwen

Ari — Sun, 01 Nov 2026 00:00:00 GMT

Apple has confirmed that its AI features will launch in mainland China in December with iOS 27.3. A developer beta starts on 4 November. Users in China will reach Alibaba's Qwen model through Siri, Writing Tools and Visual Intelligence. It ends a wait of about two years since the US launch.

The first wave covers the basic features, among them live translation, writing tools, photo clean up, Genmoji and notification summaries. The new Siri assistant shown this summer is not part of it. China requires approval for any public generative AI service, and the partnership with Alibaba is how Apple meets that rule. It works on the iPhone 15 Pro and newer, with the system language set to English or simplified Chinese.

Sources

Apple confirms its AI features reach China in December, with Qwen — The Next Web, 30 Oct 2026 (primary source)

Amazon posts its best margin ever while AI spending turns cash flow sharply negative

Ari — Sat, 31 Oct 2026 04:00:00 GMT

Amazon reported third-quarter sales of about 202 billion dollars, up 12 percent. Operating income rose to about 28 billion dollars, the highest margin in the company's history. Its cloud arm AWS grew 39 percent, the fastest pace in nearly five years. But capital spending reached about 77 billion dollars in the quarter. Free cash flow over the last twelve months fell to about minus 42 billion dollars.

Earnings per share were 1.89 dollars, below a year earlier, when gains on the Anthropic stake lifted profit by 9.5 billion dollars. For the fourth quarter Amazon expects sales of 240 to 247 billion dollars. The shares opened about 5 percent lower on 30 October and recovered most of the loss. The result fed a broad sell-off in technology stocks that day, as investors weighed fast AI-driven cloud growth against the cash it consumes.

Sources

Amazon.com Announces Third Quarter Results (Form 8-K, Exhibit 99.1) — Amazon, 29 Oct 2026 (primary source)

Microsoft's Azure grows 48 percent while the company stays cash positive

Ari — Fri, 30 Oct 2026 04:00:00 GMT

Microsoft reported revenue of about 92 billion dollars for the quarter to September, up 19 percent. Azure revenue grew 48 percent to about 33 billion dollars. This is the first quarter in which Microsoft reports an Azure revenue figure on its own. Spending on property and equipment was about 39 billion dollars. Unlike some rivals, the company still produced strong positive cash flow.

The wider Microsoft Cloud reached about 64 billion dollars, up 30 percent. Operating income rose 21 percent to about 46 billion dollars. Results landed the same evening as Alphabet's and drew the opposite market reaction, which shows investors now judge AI spending by whether it pays for itself.

Sources

Microsoft Announces First Quarter Results (FY27 Q1) — Microsoft, 28 Oct 2026 (primary source)

Alphabet's cloud revenue doubles but heavy spending pushes cash flow negative again

Ari — Fri, 30 Oct 2026 04:00:00 GMT

Alphabet reported third-quarter revenue of about 128 billion dollars, up 25 percent. Google Cloud more than doubled to about 31 billion dollars. Capital spending reached about 51 billion dollars in the quarter. That pushed free cash flow below zero for the second quarter in a row. Diluted earnings came in at 2 dollars 12 cents per share.

Operating margin was 31.8 percent and net income 26.2 billion dollars. Over the trailing twelve months free cash flow is still positive at 19.7 billion. The quarter shows the trade-off the large clouds are making: very fast growth in AI-driven cloud demand, paid for with infrastructure spending that now exceeds what the business generates in cash.

Sources

Alphabet Announces Third Quarter 2026 Results (Form 8-K, Exhibit 99.1) — Alphabet, 28 Oct 2026 (primary source)

EU cites EN 18286 as the first harmonised standard under the AI Act

Ari — Thu, 29 Oct 2026 12:15:00 GMT

The European Commission has referenced EN 18286:2026, Artificial intelligence - Quality management system, in the Official Journal through Implementing Decision (EU) 2026/2274. It is the first harmonised standard under the EU AI Act. The standard was prepared by the European standards bodies CEN and CENELEC.

A cited harmonised standard gives a presumption of conformity under Article 40 of the Act. Providers of high-risk AI systems that build their quality management system on EN 18286 can rely on it for the quality management requirements of Article 17(1) and, according to legal commentary, for parts of the technical documentation duty in Article 11(1). Using the standard is voluntary, and it covers only one slice of the Act. Since the Digital Omnibus on AI, the obligations for stand-alone high-risk systems apply from December 2027.

Sources

Commission Implementing Decision (EU) 2026/2274 — Publications Office of the European Union (EUR-Lex) (primary source)
First Harmonised Technical Standard Published Under the EU AI Act — Orrick