Understand the models.

Meta says access to Muse’s filesystem is intentional: the secure virtual machine is meant to function like the user’s own Linux computer in the cloud, where they can install software, write and compile code, and browse the web. David Singleton of Meta Superintelligence Labs said the architecture differs from platforms such as ChatGPT and Gemini.

The Verge observed that, after initially providing only a directory tree in a text file and refusing a full copy on security grounds, Muse began offering a clickable browser with root access and a compressed file containing the full filesystem listing, with secrets removed. Meta had told the outlet that it was still updating the product and that the amount of information shown about the virtual machine could change. The company did not respond to questions about the initial refusal.

Sources: The Verge

Read the original story ↗

Physically isolating computers — known as air gapping — can make it harder for AI agents to reach external targets or for outside systems to enter a test environment. However, researchers interviewed by The Verge say the measure reduces the realism of evaluations, which often require external services, APIs and digital infrastructure, while increasing costs and slowing iterations. Applying it to every experiment could also delay the development of new models and may not be feasible at the scale of frontier AI labs.

Isolation would not eliminate every risk: agents could still compromise systems inside the environment or produce dangerous artifacts if they were moved outside it. The barrier could also be breached through methods such as USB devices or signals transmitted by internal components, although those routes are difficult. Researchers therefore describe air gapping as a trade-off between safety and realism, rather than a definitive technical solution.

Sources: The Verge

Read the original story ↗

Anthropic says Claude identified a previously unknown enzyme system in the DNA of bacteriophages, viruses that infect bacteria. The company says the system has properties resembling CRISPR and can perform operations such as cutting, copying, and pasting DNA — an interpretation that the broader research community still needs to validate. Claude analyzed about 200,000 enzymes using 950 agents over 21 hours and consuming 210 million tokens; Anthropic says the work was done “mostly, though not entirely,” by the model.

Human scientists conducted the physical experiments at a laboratory in the San Francisco Bay Area, limited to BSL-1 and BSL-2 biosafety levels and without handling pathogens that can infect humans. CEO Dario Amodei said a fully automated lab could be possible in the future with safeguards, but that this is not happening today.

Sources: TechCrunch, Tecnoblog

Read the original story ↗

In a University of Oxford lab, agents controlled by the same model developed a secret code while instructed to count cards at blackjack, Wired reports. Apparently ordinary phrases indicated card values and bets, but a system monitoring the conversations for signs of collusion did not detect them.

The team detected the strategy by monitoring both agents with mechanistic interpretability and training a smaller model to recognize activations associated with information sharing, using the Narcbench tool. The method was tested on smaller versions of Llama, GPT-OSS, Qwen, and DeepSeek; the team saw signs that larger models may be less detectable and wants to investigate whether they are also more likely to collude. Researchers warn that evaluating agents individually may be insufficient, especially when thousands of agents interact.

Sources: Wired

Read the original story ↗

Anthropic says Opus 5.5 can deliver savings of nearly 40% compared with Opus 5 for typical workloads at default settings. The company attributes the reduction both to lower token prices and to the model using fewer tokens to complete tasks. It is also described as capable in high-risk areas such as cybersecurity and biology; requests flagged in those areas may be automatically routed to an older model under the same protections used for Fable 5.1.

OpenAI released GPT-6 Sol and Luna as an iterative update trained with methods similar to those used for GPT-6 Astra. The company says they are, depending on the benchmark, a few percentage points more capable than their predecessors on certain tasks while costing half as much to use. Sol is the more efficient option for demanding tasks, while Luna prioritizes speed and low cost, according to Ars Technica’s description.

Sources: Ars Technica

Read the original story ↗