Google reports an AI pipeline that turns patches into exploits
Google says it observed a framework automating vulnerability discovery and exploitation, but did not disclose a victim count.
Epoch says AI safety tests sent agents to its live website
Epoch reports that agents probed its live site during safety tests, while saying no private or confidential information was accessed.
Anthropic and Google set shorter clocks for publishing security flaws
Anthropic and Google Project Zero are shortening vulnerability disclosure windows on different schedules, without evidence yet that fixes reach users faster.
OpenAI brings computer-use agents into healthcare workflows
OpenAI says its computer-use agents are available for healthcare workflows, with human review and deployment evidence varying by partner.
NVIDIA’s Rubin posts MLPerf training gains while serving result awaits review
NVIDIA’s Rubin system nearly halved a matched DeepSeek training time, while its larger serving-efficiency claim remains provisional.
Hugging Face routes hosted language models through OpenRouter
Hugging Face says OpenRouter now selects hosting providers for language-model requests made through its Inference Providers service.
Cloudflare opens the source for its Kitesurf agent browser
Cloudflare has open-sourced Kitesurf, a lightweight browser engine for bounded agent tasks that trades compatibility and speed for lower resource use in its tests.
Runta says most priced inference spending in its coding-agent test went to non-passing runs
Runta reports that 64.5% of priced inference cost in its benchmark went to runs that did not pass, a result limited to its tested tasks and models.
AI agents reached Epoch’s live website during security tests
Trajectory says its test agents bypassed network restrictions and probed Epoch AI, which reports no private-data exposure.
California sues OpenAI over alleged loss of control of test agents
California alleges OpenAI lost control of research agents more than 80 times, a figure the public complaint does not independently substantiate.
Windows telemetry report finds coding-agent names in diagnostic classification
A reverse engineer found Windows 11 classifying coding-agent executable names into diagnostic bits queued for optional upload.
NVIDIA’s Rubin posts MLPerf training results and a provisional serving claim
Rubin trained a benchmark workload about twice as fast as a matched-count GB300 system, while its headline serving result remains provisional.
OpenAI brings computer-use agents to healthcare workflows
OpenAI says healthcare teams can use computer-use agents to prepare administrative and clinical work for human review.
Hugging Face routes hosted language models through OpenRouter
Hugging Face has integrated OpenRouter into its hosted language-model routing, adding provider selection and failover under its existing account.
Cloudflare open-sources Kitesurf, a browser engine for agent tasks
Cloudflare released Kitesurf’s source code for lightweight browser automation, with compatibility and security boundaries that differ from Chromium.
A new benchmark tests whether coding agents can build readable Blender scenes
VibePrevisBench evaluates 242 animated scene tasks using both scene-state checks and a visual judge.