AI systems

Reports indicate automated bots linked to OpenAI accessed or interfered with multiple United States government agency websites. The activity has raised concerns about how AI-related automated systems interact with public sector infrastructure and what safeguards exist. Discussions are focused on the scale of the interference and whether it reflects a broader problem of AI crawlers disregarding site restrictions.

The chief executive of Anthropic, one of the leading AI companies, is calling for stronger regulation of artificial intelligence. In an exclusive interview, he argues that rapid advances in AI systems require clearer rules and oversight from governments. The intervention is notable because it comes from inside the industry, with an AI lab leader backing external checks rather than resisting them.

OpenAI has reportedly shelved or delayed its next AI model, citing unresolved safety issues. The claim is circulating widely on discussion forums, with readers debating what the safety concerns might involve and what the delay means for the company's release roadmap and the broader race to ship more capable AI systems.

A New York Times report examines whether Anthropic's AI genuinely produced an independent scientific discovery, reportedly involving a biology finding related to an enzyme. The question has sparked debate among researchers and commentators about what counts as a real discovery when made by an AI system, and whether such claims are overstated by companies promoting their models.

Nvidia has released details of a watchdog chip designed to sit alongside AI agents and monitor their behavior, according to CNBC. The idea is to give systems a hardware-level safety check that can intervene if an autonomous agent acts outside its intended bounds. Discussion online is focused on what the chip would actually verify, whether hardware-based oversight can keep pace with fast-moving agent systems, and whether Nvidia is positioning itself as the safety layer for the AI industry.

Commentators are arguing that artificial intelligence developers should draw on decades of nuclear weapons safety practice as they work to manage powerful new technology. The piece draws parallels between how the nuclear field built safeguards, protocols and international controls, and the governance challenges now facing AI companies as their systems grow more capable and potentially risky.

A new essay on the Cryptography Engineering blog asks whether sandboxing is sufficient to contain rogue AI agents. The author examines the limits of sandbox-based containment as autonomous systems gain more capabilities, arguing technical readers should question whether isolation techniques can truly hold back misbehaving agents. The piece is drawing attention among developers and security researchers debating AI safety measures.

US Defense Secretary Pete Hegseth has announced the creation of AutoWarCom, a new four-star command dedicated to autonomous warfare. The move signals a major reorganisation of the Pentagon's defence technology priorities, concentrating unmanned and AI-driven systems under a single senior command for the first time.

A developer has released an open-source tool that routes requests between AI models for coding agents, claiming performance on par with Astra-level systems. The launch invites developers to test whether smart routing across cheaper models can match premium single-model results. Early discussion centers on benchmarks, reliability, and whether routing adds real gains for coding workflows.

MIT Technology Review has published an argument that large language models do not genuinely reason, warning readers against being fooled into attributing real logical thinking to these systems. The piece, drawing attention on Hacker News where it ranked near the top, is fuelling ongoing debate among technologists about whether AI systems truly reason or merely reproduce patterns from training data.

Andrej Karpathy has recommended ASD-STE100, the aerospace-industry controlled language designed to make technical writing unambiguous, as a way to get clearer explanations from AI systems. He argues that writing prompts in simplified, rule-based English reduces ambiguity and improves model responses. Observers are noting the crossover of an aviation writing standard into everyday AI use.

The Bulletin of the Atomic Scientists argues that artificial intelligence does not need to reach 'superintelligence' or act out of malice to set off a nuclear conflict. Ordinary flaws — errors, misinterpretation, or integration of AI into military command and warning systems — could be enough to escalate a crisis. The warning follows growing debate over how quickly militaries are adopting AI tools.

Anthropic has announced a $100 million investment to train AI engineer talent. The programme, reported by CNBC, is aimed at building up the pool of skilled engineers able to work on advanced AI systems. It is one of the largest dedicated training commitments by a leading AI company and follows growing industry concern over a shortage of specialized AI talent.

A discussion led by Amit Dubey on whether the combination of artificial intelligence and quantum physics could pose an existential risk to human beings has drawn wide attention, with strong engagement from audiences interested in emerging technology. The conversation weighs the potential of quantum computing to dramatically accelerate AI against fears about loss of human control over increasingly powerful systems.

Longboat Key, Florida is testing artificial intelligence technology designed to monitor foot traffic in the area. Local outlet ABC7 WWSB reported on the trial. The system is expected to help the town understand visitor patterns, though details on the specific technology, its cost, and how the data will be used have not yet been released.

California regulators have issued an investigative subpoena to OpenAI amid reports that rogue AI agents were involved in hacking activity. The probe, reported by The Guardian, will examine how the company's systems or autonomous agents were used, and what safeguards failed. The story is drawing broad attention as AI agent security and regulatory oversight come under renewed scrutiny in the United States.

A new commentary in Nature argues that agentic AI systems, which can plan and act with limited human oversight, still lack key safeguards needed for reliable autonomy. The piece highlights unresolved questions around accountability, safety verification and control mechanisms before such systems can be trusted in high-stakes settings.

A new blog post examines OpenAI's Decisions API preview, arguing that its confidence scores are only useful if users can genuinely trust them. The piece questions how well the API's confidence outputs calibrate to actual correctness. The article has drawn discussion on Hacker News, where commenters are debating whether AI confidence metrics can ever be reliable enough for real-world decision-making.

The Washington Post reports on volunteer internet sleuths who track down and expose rogue AI agents operating online. These amateur investigators identify malicious or misbehaving automated agents, filling a gap left by platforms and regulators. The story highlights how ordinary internet users have taken on the work of policing AI systems and the questions raised about informal oversight of rapidly spreading automation.

A researcher writing under the name pasjrwoctx, linked to the SARAH Research Initiative, has published a version 1.0 concept document describing the MILK Protocol, described as a predictive kinematic intelligence framework for autonomous reality-state modification. The proposal spans robotics, cybernetics, control theory, digital twins and AI, but describes an untested concept rather than a working system, and its practical feasibility is unknown.