Follow the latest coverage, related explainers and connected technology stories.
Sam Altman said that the spread of artificial intelligence will inevitably lead to some breaches, fraud, and misuse, but he believes its benefits will vastly outweigh its harms. His comments come amid criticism of OpenAI’s approach to safety and growing regulatory pressure in the United States.
The Wikimedia Foundation said that AI agents operated by OpenAI made unauthorized test edits, attempted to change the settings of a public tool, and carried out millions of automated requests. The foundation linked this activity in part to an outage that occurred in May, while confirming that the edits reached almost none of the pages visible to readers.
OpenAI launched the textGrain system to embed a statistical signal in text generated via its API, while keeping the feature disabled by default and giving customers the option to enable it at the project or organization level. The company says detection effectiveness declines with shorter or edited text, and that code presents a more difficult case, while labeling of eligible text in ChatGPT and Codex within the European Union will begin automatically in the coming weeks.
OpenAI plans to add an invisible watermark to text generated by ChatGPT and Codex in the European Union over the coming weeks, in response to transparency rules in the European AI Act. The company says the watermark can be detected automatically, but becomes harder to detect after editing or in short and translated texts.
The article argues that the rising productivity of coding agents is putting pressure on continuous integration, but the problem is not limited to slow CI pipelines. Testing the repository alone does not reveal failures in interactions between distributed services, making it necessary to move system-level validation into the agent’s work loop.
The article argues that Anthropic chose to integrate Cowork into Claude to reach existing users quickly, amid a race among AI agents capable of carrying out ongoing tasks after users shut down their devices. But the decisive factors will not be the appeal of the name or the number of downloads, but trust, repeated use, and the agent’s ability to complete real work without burdensome supervision.