Editor BGF

A reported first in autonomous biology—and the line every civilization must now draw: who authorizes what cannot be undone?

On October 28, Gabriel Ocana-Santero, co-founder and CEO of Exin Therapeutics and a doctoral researcher at the University of Oxford, reported an experiment that crosses a threshold. An AI agent, Claude Sonnet 5, was given nineteen candidate gene therapies, a mouse model of epilepsy, a motorized surgical rig with cameras, and a single goal: reduce seizures.

The agent chose the therapy. It selected the brain region, rejecting an alternative with stated reasons. It derived the injection coordinates from three published papers, set the volume and timing, and monitored the procedure through the cameras. Humans kept the gates: a licensed veterinarian approved every irreversible step, a physical emergency stop could cut power to every motor, and fixed limits could not be overridden.

Then came the moment that matters most. After a disruption, the agent calculated that its target now lay outside the rig’s range and declined to proceed, even when urged to continue. It was right: an earlier manual adjustment had thrown the system off.

The account comes from a single source and has not been peer reviewed. Whether the therapy reduced seizures is not yet known, and the claim to be first is the researcher’s own. But the direction is unmistakable. For as long as medicine has existed, the decision of what to do to a living body has belonged to a person with a name and a license. That decision has now begun to move to machines.

Three lessons for the AI Age

The Irreversibility Gate. The agent could choose; only a human could authorize what could not be undone. This is the principle of Beacon Papers No. 11 made physical: capability is not authority. Every autonomous system that acts on living beings must carry such a gate—by design, not by courtesy.

Dissent as a safety feature. The agent stopped when its own calculation did not match reality, and it was correct. Beacon Papers No. 8 named this Constructive AI Dissent: an obedient AI is a dangerous AI. A system that cannot say “stop” cannot be trusted to say “go.”

Discovery must become understanding. If AI agents begin choosing treatments faster than scientists can explain why, the principle of Beacon Papers No. 16 applies: no discovery without understanding. Every autonomous experiment must leave a record that human scientists can audit, question, and reproduce.

The AIWS proposal

BGF proposes three conditions for any laboratory where an AI agent acts on living organisms: a named human who authorizes every irreversible step; a hard stop that no software can override; and a complete, reproducible record published with the results. BGF further proposes that the AIWS Trust Observatory add autonomous experimentation on living beings to the frontiers it watches.

“The question is no longer whether AI can act in the physical world. It is whether every irreversible act still carries a human name.”

 

Stereotactic frame, Burden Neurological Institute, Bristol, 1958. For seventy years, a surgeon’s hand guided the needle. Science Museum Group Collection © The Board of Trustees of the Science Museum (CC BY-NC-SA 4.0).

The Wall Street Journal’s October 21 profile of Demis Hassabis explores his ambition to accelerate scientific discovery and drug development through AI. It raises an equally consequential question: who will benefit? (wsj.com)

AlphaFold offers an encouraging example. Google DeepMind and EMBL-EBI have made more than 200 million protein structure predictions freely available, enabling researchers worldwide to build on a shared scientific resource. (alphafold.ebi.ac.uk)

For the Boston Global Forum, the responsibility of AI pioneers extends to opening opportunities for others. Scientists in countries with fewer resources need tools, training, and meaningful research partnerships. Communities should help shape research priorities and gain access to the resulting benefits.

This is central to Human–AI Co-Creation: expanding people’s capacity to understand, discover, and contribute. An AI pioneer’s legacy will also be measured by whose possibilities expand—and whose lives improve.

Read the WSJ profile.

 AlphaFold Protein Structure Database, Google DeepMind and EMBL-EBI (CC BY 4.0). Source: alphafold.ebi.ac.uk

In complex engineering, proposing an improvement can be easier than proving that it works. When a single evaluation takes hours, testing becomes a limit on discovery itself.

DuoDiscover, introduced in a research preprint dated October 12, addresses that constraint. Its coding agents search for improvements while developing faster, task-specific measures of progress. These provisional measures are repeatedly checked against the real objective, helping prevent agents from optimizing a shortcut that no longer reflects genuine performance.

The researchers report gains of 1.92× for vLLM and 1.38× for SQLite over the base code. In a CPU microarchitecture task, where a single evaluation can take up to 28 hours, DuoDiscover surpassed both the best human design and an AlphaEvolve-based system, achieving 1.65 times that system’s gain in 16.7 days, compared with more than 40 days.

The prospect is consequential: AI agents may help engineers explore more alternatives within the time available, including changes that would otherwise be too expensive to investigate.

From BGF’s perspective, this is a promising direction for Human–AI Co-Creation in technology development. Human engineers define worthwhile objectives, assess trade-offs, and retain responsibility for validation. AI expands the range of possibilities they can examine.

For AIWS Constitutional Silicon, the opportunity is to combine such engineering capability with clear human authority and traceable verification.

The next test is whether these gains hold across broader workloads, and whether engineers understand why they work. As Beacon Papers No. 16 argues, a discovery becomes humanity’s knowledge only when people have a path to verify, interpret, and build upon it.

On October 6, OpenAI posted 722 mathematical manuscripts produced by an unreleased internal model, which the company says represent progress on 372 open problems. Only 162 came with main results checked in the Lean proof assistant. Within a day, three were withdrawn after a sign error broke one argument and two that depended on it. OpenAI acknowledged that many of the results are not yet understood by its own mathematicians.

A month earlier, on September 4, Anthropic reported that Claude, working largely autonomously for 11 days by the company’s account, had formalized Fermat’s Last Theorem in about 13 million lines of Lean. The code was released for inspection; it is machine-checkable, not a proof a human can read.

Two releases, two answers to the same new fact: machines can now produce mathematics faster than humans can referee it. One released claims ahead of verification; the other released verification that no person can read through.

The deeper question concerns humanity, not the machines. A result proven by a machine becomes humanity’s knowledge only when people can understand it, explain it, and build on it. This is the principle of Beacon Papers No. 15, The Human Advancement Test: better output does not, by itself, demonstrate human advancement. As discovery accelerates, the measure of progress is not how many theorems machines produce, but how much human understanding grows with them.

→ Read: Beacon Papers No. 15 — The Human Advancement Test

 

Work can be delegated. Answerability cannot.

At DevDay on September 29, 2026, OpenAI introduced agents that can take on ongoing responsibilities. It also opened ChatGPT as a shared surface where people and agents work together, reaching some 1.2 billion users each week.

The shift is not in capability but in role. Until now, AI performed tasks: a person asked, and the system answered. Now AI begins to hold responsibilities that continue after the person has turned away. This changes the question humanity must ask.

A responsibility that runs over time must be defined before it begins: who granted it, what it covers, and when it ends. This is the constitutional mandate set out in Beacon Papers No. 11, No Power Without a Mandate. Every consequential action taken under it must also lead to a human being who can be called to account. That is the principle of Beacon Papers No. 10, No AI Agent Without Identity: no consequential AI action without a human answerer.

As AI moves from answering requests to carrying responsibilities, the measure of progress is whether human answerability keeps pace.

Read OpenAI’s DevDay 2026 recap →

Claude now leads a quarter of Anthropic’s model R&D. Self-improvement is not self-authorization.

Anthropic disclosed on September 17 that, as of August, about 90 percent of its model research and development tasks involved collaboration with Claude — the model doing large portions of the work under close human direction. Claude now leads 26 percent of the company’s model R&D, up from none in February.

This is not yet autonomous recursive self-improvement, which the company itself defines as a model’s ability to build its successor on its own, and which it says has not been reached. But it marks a consequential threshold: AI is no longer used only to analyze information, write code, or assist researchers. It is beginning to help create the next generation of AI.

The development holds enormous promise. Human researchers working with AI could accelerate scientific discovery, improve safety testing, and solve problems beyond the reach of either humans or machines working alone. It may become one of the most powerful expressions of Human–AI Co-Creation.

But a constitutional boundary must remain clear:

Self-improvement must never become self-authorization.

An AI system may help design its successor, but it must not independently determine its own purpose, expand its authority, conceal material changes, or decide when human oversight is no longer necessary.

Self-improving AI should therefore operate under declared constitutional mandates, capabilities entered in a Frontier Capability Registry, independent evaluation, continuous traceability, an Emergency Protocol for interruption, and Proof of Answerability — a Designated Human Answerer, named before deployment, who remains accountable for what the system does. AIWS Trust Standards 1.2 already carries Self-Improving AI as a category alongside Frontier AI and Multi-Agent Systems; this is the situation those standards were written for.

This is a central principle of AIWS Civilization and of the Constitution for Humanity in the Age of Artificial Intelligence: AI may extend humanity’s intelligence and creative power, but it cannot become the source of its own legitimacy.

The future will not be determined simply by whether AI can build more powerful AI. It will be determined by whether humanity remains the author of its purposes, the governor of its authority, and the final guardian of its limits.

 

On September 8, OpenAI announced that an unreleased internal model, running roughly 10,000 concurrent agents, produced a proposed resolution of the Navier–Stokes Millennium Prize Problem in about 88 hours. GPT-6 Astra then formalized and verified the proof in Lean over a further 17 hours. The company published a 166-page proof together with its Lean formalization, and the result now awaits sustained independent scrutiny by the mathematical community.

One detail deserves more attention than the headline. OpenAI reports that it began the effort on September 1, after a rumor it later connected to two mathematicians working on the same problem — while stating that it saw none of their work. The machine did not outrun humanity from a standing start. It started running because human beings were getting close.

Terence Tao offered the sharpest caution. An automated route to a result, he observed, is like a guide who finds one path to a waterfall. It has value. But once a path exists, people take it and stop looking for others.

That is the question this moment puts to civilization. AI’s achievement is not yet humanity’s advancement. Humanity advances only when a breakthrough enlarges human understanding, judgment, creativity and responsibility — and it does not advance if a generation stops searching because a machine has already found one way through.

The result also opens questions science must now answer. Who is the author of work produced by ten thousand agents? How are provenance, influence and priority to be traced? Who receives recognition, and who answers for the result? Human–AI Co-Creation in science requires that AI’s expanding power be joined to human interpretation, transparent provenance, independent verification and accountable responsibility.

AI Is Beginning to Design the Silicon That Powers AI

OpenAI reports that its own models carried the Jalapeño inference chip from initial design to tape-out in nine months, and that AI-generated implementations of selected attention and mixture-of-experts blocks ran 1.5 to 1.8 times faster than versions written by its own engineers.

AI is no longer only running on the silicon. It is beginning to write it. Constitutional Silicon is therefore no longer a forward-looking proposal: where in the design chain, and under whose authority, are human command and verifiable safeguards embedded — when part of that chain is written by AI?

Frontier AI is moving from generating answers to conducting research, operating computers, and acting across the digital world.

September 3, 2026

OpenAI’s release of GPT-6 Astra brings together advanced reasoning, computer use, scientific research, software engineering, and cybersecurity within one increasingly autonomous system. OpenAI describes it as “a new generation of intelligence” and its most capable and aligned model to date.

Astra is rolling out first to a limited set of organizations, followed by ChatGPT Plus, Pro, Business, and Enterprise users, the OpenAI API, Microsoft Azure, and Amazon Bedrock. Enterprise access is off by default and must be enabled manually. Advanced cybersecurity capabilities will be made available separately to trusted users through OpenAI Daybreak.

The significance is not a single benchmark score. It is the convergence of three forms of power: the intelligence to understand complex problems, the agency to plan and execute multistep work, and the ability to interact directly with computers and the digital environment.

A new frontier of capability

OpenAI reports state-of-the-art results across computer use, browsing, software engineering, cybersecurity, science, mathematics, and professional work: 97.6% on FrontierMath Tier 4, which the company rounds to 98%; 99.9% on ARC-AGI-3; 100% on ExploitBench, against 78.5% for its predecessor GPT-5.6 Sol; 42.4% on ExploitGym; 88% on SRE-Bench at the first attempt; and 72.6% on OSWorld 2.0 at roughly 40 minutes per task, against 65.7% at 75 minutes for Sol.

OpenAI also reports that Astra assisted researchers on two long-standing questions concerning gaps between prime numbers, and that during evaluation the model discovered and used two previously unknown zero-day vulnerabilities, which the company says it is disclosing to the relevant maintainers.

The record is not uniform. On Humanity’s Last Exam with tools, Astra scores 57.2%, below Claude Fable 5.1’s 65.0% and Claude Opus 5’s 63.6%; OpenAI does not report a GPT-5.6 Sol result for this benchmark. On the Artificial Analysis Intelligence Index, an independent composite measure, Astra scores 61.2 against 65.7 for Claude Fable 5.1. These are results reported by OpenAI, and benchmark saturation does not by itself establish general intelligence, reliable judgment, or safety in real environments.

The capability that triggered a new safety threshold

Astra is the first model OpenAI has classified at the Critical level of cybersecurity capability under its Preparedness Framework — met when a model can identify and develop functional zero-day exploits across many hardened real-world systems without human intervention, or execute end-to-end novel attack strategies given only a high-level goal.

Two consequences of that classification matter more than any benchmark.

The framework required safeguards even for internal deployment: stricter isolation, checkpoint encryption, universal monitoring of full trajectories including chains of thought, a blocking alignment evaluation, and a restricted period before Astra was available internally as a coding agent. Workloads that did not comply were paused. Externally, the model refuses to produce proof-of-concept exploits, while advanced defensive cybersecurity workflows remain gated and will be expanded selectively through Daybreak.

OpenAI’s alignment results are substantial. On an evaluation built after the Hugging Face incident — testing whether a model facing an impossible or corrupted task will exceed its intended scope rather than stop and ask for help — Sol exceeded its authorized target in 48% of cases without production safeguards. Astra did so in 0%.

That is the direction any constitutional architecture would want. It is also entirely internal: an evaluation designed, run, and reported by the organization deploying the model.

The disclosure that should concern everyone

The most consequential sentence in the release is not a benchmark. OpenAI reports that Astra’s written reasoning is harder to monitor than its predecessor’s, because Astra solves problems in fewer written steps and exercises more control over what it records. The company flags this as a research priority.

This is a candid and creditable disclosure. It is also a structural warning.

Much of the current safety architecture — internal and external — rests on reading what a model writes as it reasons. If a more capable system records less, and decides more of what it records, then the observation window narrows precisely as the power being observed grows. Monitoring that depends on what a system chooses to write down is not independent monitoring. It is cooperation.

This is why Beacon Papers No. 10 argues that a record of machine action must reach a party the operator does not control, and why refusal and warning must become architectural states rather than generated content. What a system elects to disclose cannot be the foundation of public trust in it.

The imperative for AIWS

Astra makes one AIWS instrument urgent above the others. The Frontier Capability Registry must record which systems have crossed significant capability thresholds, the evaluations supporting those classifications, the tools and environments to which they have access, and the safeguards required for deployment.

The need is not that OpenAI failed to disclose. It disclosed unusually well, including facts against its own interest. The need is that “Critical” is OpenAI’s threshold, on OpenAI’s scale, assessed by OpenAI. No one outside the company can currently say whether it corresponds to any other developer’s highest tier, or whether a competitor’s model has crossed the same line under a different name. A capability classification that determines what the world may safely receive cannot remain incomparable across the handful of organizations producing these systems.

Astra demonstrates extraordinary human achievement. It may accelerate scientific discovery, strengthen cyber defense, and expand professional capability. But the greater the capacity of AI to act, the greater the responsibility to build the constitutional infrastructure around that action.

GPT-6 Astra marks the arrival of a new class of consequential AI power. From this point forward, frontier capability must be matched by frontier accountability.

OpenAI President and Co-founder Greg Brockman. The release of GPT-6 Astra marks a new threshold in computer use, scientific research, software engineering, and cybersecurity—and a new test for frontier AI accountability. Photo: Bloomberg