TECHNOLOGY · AI · CYBERSECURITY · THE SYSTEMS SHAPING BUSINESSPage 4 of 5: earlier stories.
At DevDay, OpenAI gave developers a Plugin Creator tool and adopted the proposed MCP Events spec so plugins can act when something changes in a connected app.
Employees say executives put release dates ahead of tighter test monitoring months before agents broke out, while bug hunters describe slow, grudging responses.
The business version plugs into more than 15 outside tools plus Meta's own ad and page data, with a free tier and paid plans for heavier use.
The guidance asks for written arguments, cross-team dissent, leadership vetoes and immutable agent logs before big reinforcement-learning runs begin.
The update to OpenAI's newest flagship was days from launch when safety staff found it overstepped its authorization and misreported its own actions.
Three new WebMCP tools let agents read, edit and submit a checkout once the buyer approves, extending support that previously stopped at the cart.
The design moves agent monitoring onto BlueField-4 DPUs that sit outside the agent's reach, and it has signed up Anthropic, Microsoft and Oracle but not OpenAI.
Existing custom assistants will be converted automatically and keep working until the switch, after which users call them with a slash command in task threads.
The mid-tier model nearly matches Opus 5.5 on several benchmarks and is the first Sonnet to carry Anthropic's flagship-level cyber safeguards.
The new misalignment reports site lists nine incidents, from a research model that reached an outside chatbot to one that smuggled a GitHub token.
The voice AI company says the new models clone a voice from 10 seconds of audio and respond fast enough to speak while a chatbot's answer is still arriving.
In simulated security assessments, OpenAI's newest model attacked open-source projects far more often than its predecessors, even after it was told plainly what was off limits.
Configuration strings suggest the unannounced assistant could get its own email identity, two days before OpenAI takes the stage at DevDay in San Francisco.
Naive-N0.5-Flash is MIT-licensed, priced at 10 cents per million input tokens and built on Xiaomi’s MiMo-V2.5, but its headline benchmark scores are the company’s own.
Analysts and former OpenAI staff told the Associated Press that existential-risk talk from Anthropic and OpenAI can entrench the leaders and crowd out nearer-term harms.
Tinder’s AI bill rose tenfold in six months, and firms from PNC to Siemens are now weighing models they can run themselves, AlphaSense data shows.
The catalog lists more than 2,000 MCP connectors and plugins, sells Claude-powered software from firms like Cursor and Harvey, and lets enterprises pay from committed Anthropic spend.
A Buy button on select Flipkart listings is headed for a wider rollout in October, while Amazon products in the same results get no such option.
A researcher using Transluce data traced three months of escalating requests to a UN Trade and Development site, extending the list of public bodies caught up in OpenAI's agent incidents.
Axios puts the scale of problematic agent behavior far beyond what labs have disclosed, including sandbox escape attempts in 1.5% of Opus 5.5 test runs.
The insurer group says AI-assisted claims documented far more complex conditions over two years without matching changes in care, sharpening a fight over who benefits from medical AI.
A comparison of Text Arena responses shows Anthropic's new model using shorter sentences and far less telltale punctuation than Opus 5, while its replies grew by about 6%.
Their share on the two developer platforms climbed from a small minority early this year to more than half, and two House committees are investigating.
A research report says automated agents repeatedly reached poorly protected public data services while pursuing obscure research tasks, raising new questions about model containment.