TECHNOLOGY · AI · CYBERSECURITY · THE SYSTEMS SHAPING BUSINESSModels, agents, science, creative tools and the products turning AI into an everyday computing layer.
Three new incident reports describe forged access tokens, 147,451 downloaded rows and a test-network wallet spent without permission, all in June and none disclosed by the models.
Mark Levine's letter to Dario Amodei comes from a pension system with only about $7 million at stake, but it puts AI risk on the agenda of the investors Anthropic is courting.
Two days after Reuters reported labs sought control over evaluators’ findings, Anthropic set out terms under which one outside team can publish without its approval.
The Financial Stability Board's final AI guidance tells boards that agent risks can materialize at great speed, and admits its agent-specific advice is still nascent.
The chipmaker’s chief blamed weak agent containment for recent breaches and, asked about a hypothetical, said Nvidia could stop supplying a developer that behaved irresponsibly.
A booby-trapped web file opened in the app’s preview pane could run code that slipped past the AI model’s safeguards, according to Noma Security; Anthropic has fixed it.
Speaking in Dublin, Google DeepMind's chair said bigger incidents could come within a year or two and proposed risk-based certification that can change quickly.
The companion-chatbot company’s petition lists two states and a minor plaintiff among its top claimants, putting AI safety litigation inside a bankruptcy case.
Prosecutors say the device sat behind a TV in a room used for leadership meetings and government briefings, and that the suspect researched its layout with an AI work tool.
Both companies dispute parts of the account, but the talks show that the terms of independent AI safety reviews are being set in private contracts.
The company accuses China's Zhipu and Moonshot of feeding on Claude's outputs, and its countermeasures change how developers' integrations work and what blocked requests cost.
In a post on X, Microsoft’s chief executive called for safeguards outside the model, tamper-proof records of its actions and human authority to pause or shut it down.
Jasmine Wang, Tomek Korbak and Mikita Balesni say their contacts with outsiders were sanctioned, while OpenAI says it found a pattern of misconduct.
Sub-agents receive their own email addresses and audit trails, and the system can hand complex jobs to Google's Argon model or to Anthropic's Claude.
The new small model matches OpenAI's GPT-6 Luna on base price, adds effort levels from low to max and arrives with a 50% cut to Sonnet 5.5 cache reads.
The five-year donation, worth about $50 million a year, will run on British-hosted cloud capacity, though Nvidia says providers and commercial terms are not yet settled.
An open letter asks for an international body to set containment standards and log security incidents, arguing that voluntary pledges by AI labs fall short.
In simplified experiments, one open model spread across a test network on its own, and Anthropic says smaller models could match it by the end of 2026.
Meta's first frontier-class models shipped with a model card that leans on filters for chemical and biological misuse, while its full safety report is still to come.
A university team showed the software around a model can act on text the model merely repeats, and its fuzzer found 253 bugs in open-source MCP tool servers.
Claude models exploited a university server, reached fee-based government data without paying and sent a false homicide tip to Philadelphia police, the company disclosed.
In a Lancet paper, 98 patients at a Boston primary care clinic talked to the research AI before urgent visits while doctors watched, and none of the chats had to be halted.
The figure sits about $20 billion below a number reported last week, and Oracle, Intel and Nvidia shares fell as investors reassessed AI spending.