6 November 2026 · Security, research and health

A model to find vulnerabilities, an assessment of how AI helps develop more AI, and a study using smartwatch sensors. Today’s briefing selects three publications from 5 November. Their common thread is the relationship between what systems can do and how their results can be checked.

OpenAI · A model to hunt for flaws

OpenAI has introduced GPT-6 Cyber, a model dedicated to cybersecurity. The company reports that a research campaign identified 81 previously unknown high-severity vulnerabilities in Firefox, disclosed to Mozilla with proposed fixes. Access is restricted through Daybreak Red: the model is designed to stay within authorized scope. Published results and safety evaluations do not guarantee that errors cannot occur.

OpenAI · GPT-6 Cyber, 5 November 2026 ↗

METR · Research accelerates, but much of the evidence is not public

METR estimates that Anthropic is improving model capabilities on well-defined tasks at roughly twice the rate of the historical period examined. The assessment, published on 5 November, carries substantial uncertainty. Its authors believe AI assistance in research probably accounts for much of the acceleration, but note that Anthropic redacted much of the most informative evidence. The public therefore cannot directly examine all the evidence behind the conclusion.

METR · Assessment of Anthropic’s AI research, 5 November 2026 ↗

Google Research · Hypertension signals from the wrist

Google Research describes a system combining a smartwatch’s optical and motion sensors to identify signs of hypertension across several days. Clinical validation involved 196 participants at five US sites, compared against 24-hour ambulatory blood pressure monitoring. The authors present it as exploratory screening research: it does not replace a clinical arm cuff and does not provide a standalone medical diagnosis.

Google Research · Wearable sensors and hypertension, 5 November 2026 ↗

The desk’s view · Capabilities and evidence must grow together

In the desk’s view, the question is not only how much work can be delegated to AI. It also matters who can inspect the evidence, repeat the tests and decide when a result still needs human verification.

Sources and documents

Document date or reference period: Selection for 6 November 2026 · Original sources published on 5 November 2026.

Desk selection and summaries with AI assistance, based on the linked original sources, consulted on 6 November 2026. This briefing covers publications from 5 November, rather than only news released during the Italian night. Company results and research assessments remain attributed to their respective sources.

Report an update or correction →