AI Morning Brief — Mon, Sep 28
Top story
Nvidia launches an open platform for controlling AI agents
Nvidia's Open Agent Safety Platform combines the broadly available OpenShell runtime with Sentry, a hardware-monitoring reference design for BlueField-4. Nvidia says Sentry can quarantine agents that stray beyond their assigned boundaries; CNBC independently covered the launch and interviewed a company executive, but real-world effectiveness has not been established. ·
Why it mattersDevelopers can put access limits outside an agent's own model and application logic, where recent breakouts have exposed weaknesses.
Also
State AI reporting laws may leave agent breakouts undisclosed
MIT Technology Review interviewed legal experts who say recent incidents may fall below several states' mandatory “critical safety incident” thresholds, leaving investigators to use other legal powers.
Why it mattersPublic disclosure and accountability can hinge on damage thresholds rather than the loss of agent control itself.
Tools & releases
H releases two open-weight Holo4 computer-use models
H made 27B and 35B-A3B models available through its API and for download; it says they can work through screens, code and APIs, and its own OSWorld 2.0 results remain vendor benchmarks.
Why it mattersDevelopers can test a downloadable agent model across several software interfaces.