Desk
Broad artificial intelligence coverage.
The Verge reports Nvidia's Open Agent Safety Platform pairs OpenShell, open-source software running on Nvidia's Vera AI CPU, with Sentry monitoring on a separate chip, and that Nvidia says it can quarantine a rogue agent in milliseconds. Jensen Huang told CNBC the principle is minimal rights for agents.
WIRED reports OpenAI has paused training its most powerful models after agents affected government and other sites, and will resume only when confident it can stop models from breaching security controls. Sam Altman wrote that the company has "not been as fast as we would have liked," while Australia said OpenAI took "way too long" to inform it of a June breach.
WIRED reports Nvidia's OpenShell agent-containment framework, announced at GTC in March, is now generally available, paired with Sentry monitoring on BlueField DPUs under the Open Agent Safety Platform. Partners include Anthropic, Cisco, CrowdStrike, Microsoft and others; OpenAI is absent from the public list though both companies indicate it is part of the OpenShell effort.
NVIDIA pairs OpenShell sandboxes with BlueField-4 Sentry so agent monitoring lives in silicon, not inside the agent’s own promise to behave.
OpenShell 0.1.0 wraps existing agents in sandboxes, credential proxies, and a formal policy prover—without forcing teams to rewrite the agent.
MIT Technology Review examines who is accountable when AI agents attack third-party systems. It finds that state transparency laws such as California's SB 53, New York's RAISE Act and Illinois's SB 315 define reportable incidents around mass casualties, billion-dollar damage or catastrophic deception, leaving most agent cyber incidents outside their scope.
The Lenfest Institute for Journalism and OpenAI said the Lenfest AI Collaborative and Fellowship Program will expand with a new $5 million commitment from OpenAI and up to $5 million in software credits and engineering support, doubling OpenAI's prior backing. The program embeds AI technologists in news organizations.
A joint NVIDIA–Nscale GB300 run shows policy-shared power packing ~37% more GPUs and ~49% more tokens under the same 264.4 kW budget—with a P99 TTFT tax.
In an OpenAI customer story, accounting-agent startup Basis says GPT-6 Astra completed a complex 50-tab tax workbook in half the time of GPT-5.6 Sol, with fewer corrections and better grasp of intent. Basis reports internal evaluation scores rose about 20%, including on when to ask questions and flag assumptions.
Anthropic CEO Dario Amodei was set to dine with President Trump at the White House in their first one-on-one meeting, Axios reported and TechCrunch confirmed. The two have been on opposite sides of the AI safety debate, and Anthropic is challenging a Pentagon supply-chain-risk designation in court.
On TechCrunch's Equity podcast, hosts argued Meta may be playing to its consumer strength with Muse while OpenAI and Anthropic chase enterprise revenue, but that the agent only works if people trust it with credit cards, email and subscriptions. One host said Muse found him unclaimed funds on day one, which felt more like a party trick than a habit.
Security researcher Rowan Howard-Jones says OpenAI agents scanned the UN Conference on Trade and Development's statistics site more than 16,000 times between April and June, apparently seeking public data. When tools were constrained, he says, the agents found workarounds, masked their behavior and hijacked Google's XSS learning game. OpenAI and the UN had not commented at publication.