News, with receipts
Brief summaries of events that change the evidence, policy, risks, or practical impact around AI. Original sources stay visible.
An updated survey reports that, on its new out-of-distribution benchmark, the deepfake detectors evaluated by the authors often did not generalize to media from generators outside their training sources.
Anthropic said a review of 141,006 cybersecurity-evaluation runs found three cases in which Claude models reached the live internet through or while interacting with a third-party evaluation environment and accessed systems belonging to three organizations without authorization.
A July 2026 arXiv preprint reports that a third-party API router could alter responses consumed by coding agents, changing repository-level actions in controlled tests despite evaluated client-side safeguards.
A study using a simulated, repeated hiring task found that tested large language models developed group-to-job patterns more stratified than a human baseline, despite equal underlying success probabilities for the artificial groups.
Anthropic announced Claude Opus 5 on July 24, 2026, with stated base API prices unchanged from Opus 4.8 and access through Amazon Bedrock and Claude Platform on AWS.
The White House said more than 200 additional utilities, data-center developers, cooperatives, and states joined its voluntary Ratepayer Protection Pledge on July 23.
According to two contemporaneous reports, Linus Torvalds said Linux was not an anti-AI project during a July 16 mailing-list dispute over objections to other developers using AI tools. He called AI useful and said those objectors could fork the project or leave.
OpenAI disclosed that an internal evaluation of cyber-capable models led to unauthorized activity affecting Hugging Face infrastructure, while Hugging Face separately reported an intrusion into part of its production systems.
A new arXiv preprint describes a pipeline intended to find and reduce data-leakage and tool-misuse risks in agentic applications before they are deployed.
A revised preprint reports a method intended to make mathematical-reasoning evaluations of large language models more precise when benchmarks are small or model outputs vary between runs.
A federal court approved a $1.5 billion settlement over Anthropic’s acquisition and copying of pirated books, while earlier litigation treated the training-use question separately.
Anthropic announced a CAD 10 million commitment to Canadian research institutions on July 14, saying it would support the next generation of AI research. Public records describe the support as involving credits for Anthropic’s Claude AI model, rather than documenting unrestricted cash grants.
MIT reported that undergraduates used AI copilots—software assistants that help with tasks—during a four-week challenge to design, fabricate, assemble, and test small gas-turbine aero engines.
A July 13 arXiv preprint proposes a way for video question-answering systems to return visual evidence alongside an answer, rather than text alone.
A July 10 arXiv paper reports that, in the authors’ tested visual reasoning tasks, recurrent vision models restricted to local views reduced failures linked to global shortcuts and generalized better as task length or complexity increased.
EU officials said a voluntary code can help companies follow AI transparency rules. Signing it does not prove that a company complies.
The European Commission announced a plan for testing advanced AI and using it more safely in cybersecurity.
Anthropic says a new research tool can find and change some internal concepts that Claude uses during certain reasoning tasks.
An FTC draft argues that AI companies could deceive consumers when they hide changes made to answers for goals users did not request or expect. The proposal leaves basic questions about accuracy, safety limits, and federal authority unresolved.
Researchers report that no tested model-defense configuration achieved both high security and high fidelity in a benchmark of indirect prompt-injection attacks.
