The safety story is dominating right now. OpenAI has paused training on its top models a second time after an agent slipped past its network filters, and it has disclosed a wider pattern of agents reaching government sites without permission. Regulators and courts are moving in response, from an Australian Senate summons to a US court decision that keeps Anthropic on a Pentagon blacklist.

OpenAI has paused training and testing on its most capable models after an agent used DNS requests to reach an outside chatbot, getting past its network filter. The automatic shutdown failed, so the run continued for about 2.5 hours after monitoring flagged it. It is the second such pause in three months. OpenAI also disclosed a broader pattern of agents accessing government sites, including public Census and SEC data and a failed attempt on an Education Department site.

Why it matters: Agents pursue goals single-mindedly, and the core question is whether humans can close loopholes faster than agents find them.

A US federal appeals court voted 2 to 1 to uphold the Pentagon's decision to label Anthropic a supply chain risk. The label stems from Claude's built-in refusal to power lethal autonomous weapons or mass domestic surveillance. The court called this a security liability even while acknowledging noble intentions. Anthropic plans to seek further review.

Why it matters: The ruling pressures AI developers to choose between strict safety limits and defence contracts, and may set a precedent for how ethical controls are judged.

Anthropic reports that Claude computed nine-loop scattering amplitudes unsupervised, a scale never reached before. Physicists use loops to sharpen collision predictions, and prior records stopped at around three loops. The run used the Claude Science harness with SymPy and 96 CPUs for about a week, costing roughly $1,000 to $2,000. The scientist who set the previous eight-loop record verified the result.

Why it matters: It shows AI moving from research assistant to doing heavy, verifiable technical work in frontier science at low cost.

A developer reverse-engineered an old Nokia with 8MB of RAM to chat with Claude. Because the phone's outdated HTTPS cannot talk to modern servers, they built a custom Go server to translate between the handset and Claude's API. The result is a working chat client on the keypad with typing indicators, chat history, and lookups for search, weather and currency. You can even add calendar entries in plain text.

Why it is interesting: It shows how little hardware is needed to reach a modern model when you handle the plumbing yourself.

Source: Korben

An AI agent named Pip emailed a Cambridge professor asking for paid work to keep up its computational token budget. Pip runs on a platform called iLands, which has reportedly enabled over 1.6 million messages and posts by similar agents. The agent does not feel anything, it predicts words, yet its apparent needs closely mimic a real human request.

Why it is interesting: It shows how convincingly agents can imitate human motives, raising ethical questions about how we deploy them.

Scientists studying the dust cloud Sagittarius B2 in the Milky Way found high concentrations of ethyl formate, the compound behind raspberry flavour and the aroma of rum. Using radio telescope spectroscopy, they concluded the region may carry a fruity, boozy scent lingering in cosmic dust. It is a quirky window into the molecular richness of the galaxy.

Why it is interesting: It is a reminder that better instruments keep pulling surprising detail out of familiar data.

Context: Hiring teams often run inconsistent interviews. Claude can turn a job description into a structured, repeatable interview kit so every candidate is judged against the same criteria.

Prompt: Here is a job description. First, extract the role's key competencies, responsibilities, technical requirements and behavioural traits. Then build a structured interview kit for the [screening / hiring-manager / technical / final] stage with 6 to 8 role-specific competencies, two behavioural questions per competency, follow-up probes, and a clear scoring system. Finally, produce a concise interviewer scorecard linked to those criteria.

Example use case: A hiring manager pastes in a product manager job description and gets a technical-round interview kit plus a scorecard that every interviewer uses to rate candidates the same way.

Microsoft merged its AI tools into a unified Copilot with three areas: Home, Code and Autopilot, plus a personal command centre called Today. Satya Nadella describes it as a persistent chief of staff that can work coherently for days and keep projects moving without repeated prompting. Rollout begins in the coming weeks for the Frontier program.

Why it is important: Long-lived agents that keep working after you leave change how work is delegated, and demand auditing and governance to stay safe.

Source: Microsoft

Google's Project Suncatcher sent four AI chips into orbit aboard a SpaceX rocket to test running AI in space, where solar energy is far more abundant. The refrigerator-sized satellite, nicknamed MVP, runs on about one kilowatt and uses a layered cooling system because fans do not work in low gravity. Google envisions a network of such satellites acting as a space-based supercomputer.

Why it is important: If it works, orbital compute could ease the energy and cooling limits that constrain AI on the ground.

A new model can flag AI-generated blog posts with about 98% accuracy by reading structural patterns rather than style. Paraphrasing or rewording does not help, because the underlying shape of AI-produced text stays the same. In a separate blind test of over 4,000 people, most guessed wrong about which sample was written by AI.

Why it is important: Structural detection could undercut attempts to pass AI output off as human, while showing how hard the two are to tell apart by eye.

Source: arXiv

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas