OpenAI · Dots · GPT-6.1 Sol
OpenAI shelved an agent model Monday and launched an agent product Tuesday
The timing is almost comically instructive. GPT-6.1 Astra’s planned release was canceled after safety regressions; the next day, OpenAI introduced Dots and GPT-6.1 Sol. Dots use the existing GPT-6 Astra, with their own cloud computers, connected tools and the ability to pursue recurring work. Sol is a separate model release, not the withheld Astra successor.
The important change is responsibility that persists. A proposal changes when a customer changes requirements. A software project needs another update. Research keeps generating new information. A chatbot waits for the next question; a persistent agent can keep track of what still needs doing. That turns permissions from a single decision into something that must survive a changing task.
OpenAI’s published rules include custom boundaries, action review and safety monitoring. The most sensitive actions, such as password changes, require a human takeover; permanent deletion can require approval each time. These controls deserve scrutiny alongside capability claims. The race is toward the most capable agent you can actually trust with the keys.
Source: OpenAI system card ↗Source: OpenAI permission rules ↗Source: GPT-6.1 Sol announcement ↗
Visa · Cybersecurity · Agentic defense
Visa would prefer the defensive AI to arrive before the offensive AI
Fresh reporting puts Visa’s AI defenses in the spotlight, but the open-source project has a longer history: the company’s announcement dates to June 10 and has since been updated. Its Visa Vulnerability Agentic Harness is a reference implementation for vulnerability discovery, remediation and validation, informed by participation in Anthropic’s Project Glasswing.
Visa describes a control plane for autonomous detection, defense, remediation and containment, with human accountability at critical decisions. The logic is straightforward: if attackers can automate increasingly long chains of work, a defender waiting for a person to inspect every alert may operate at the wrong timescale. Releasing a harness gives other security teams something concrete to inspect and adapt; it does not establish universal protection.
Payment networks make the authority problem especially vivid. Who authorized a transaction? Did a purchasing agent exceed its spending boundary? Did another system manipulate it? Those are security questions and questions about machine authority. Machines may attack, defend and transact at machine speed. Humans still have to reconcile the statement.
Source: Visa ↗
New Mexico · Quantum computing · Infrastructure
Quantum computers got a building designed around their dislike of everything
Roadrunner Quantum Lab opened September 29 in Albuquerque’s Innovation District. The 35,000-square-foot facility is part of New Mexico’s broader $450 million quantum commitment, providing laboratory space, specialized equipment and collaboration among companies, national laboratories and researchers.
Its announced participants span several approaches to quantum hardware and control, including IonQ, QuEra, Diraq and Quantum Machines, alongside Sandia, Los Alamos and the Quantum New Mexico Institute. Shared infrastructure can help companies turn an isolated experiment into a repeatable system without rebuilding every supporting resource themselves.
Quantum computing’s impressive promises rest on decidedly physical requirements: lasers, electronics, refrigeration, vacuum and isolation, depending on the hardware. A qubit count tells only part of the engineering story. Classical computing conquered ordinary rooms; quantum computing is still negotiating the terms of its tenancy. The building is becoming part of the computer.
Source: Roadrunner Venture Studios ↗
