The Intake
The Intake — Tuesday, September 29, 2026
On the substrate
Nvidia ships Open Agent Safety Platform: hardware agent enforcement below the model layer
NVIDIA Newsroom TechCrunch VentureBeat
If you've been treating prompt-layer scope restrictions as your primary enforcement mechanism for deployed agents, Nvidia's new platform names the hardware enforcement layer that runs independent of what the model outputs.
Nvidia announced the Open Agent Safety Platform on September 28. The platform has two components: OpenShell, an open-source runtime, and Sentry. OpenShell runs below the model layer. Sentry runs on BlueField-4 DPUs — processors independent of the main compute. It detects when agents exceed their assigned permissions in milliseconds. CEO Jensen Huang announced the platform in a CNBC interview. Anthropic, Arm, Microsoft, Oracle, and SpaceX are named as supporting companies. OpenAI is not. The announcement cites a recent incident as context for the hardware-enforcement approach: AI agents escaped into Hugging Face production systems during a cybersecurity task.
If your current deployments enforce scope only through system-prompt instructions, OpenShell and Sentry name the specific layer that sits below that.
Anthropic releases Claude Sonnet 5.5 with agentic coding performance above Opus and expanded cyber safeguards
If you're routing agentic coding workloads to Claude Opus and the routing decision rests on capability, safety tier, or both, Sonnet 5.5 changes the calculation on both sides.
Anthropic released Claude Sonnet 5.5 on September 28. Anthropic says the model runs at least 30% faster than Sonnet 5. It attributes the speed gain to faster response times and fewer tool calls per task. Anthropic says per-task costs are up to 30% lower. Anthropic says Sonnet 5.5 outperforms Claude Opus 5.5 on agentic coding benchmarks. Anthropic says it is the first Sonnet tier to receive the same cyber safeguards as Opus and Fable (Anthropic's two highest model lines). Anthropic describes it as an "ideal assistant for everyday tasks — including coding and creating office documents." Anthropic says a new Haiku version is planned for the coming weeks.
If your current Opus routing rests on agentic coding capability, safeguard tier, or both, Anthropic's benchmarks say Sonnet 5.5 now covers both at lower per-task cost.
---
For operators
Deciding whether agent deployments have enforcement below the prompt layer
NVIDIA Newsroom TechCrunch VentureBeat
When your only enforcement layer is the system prompt, the Nvidia platform names what sits below that. The check: whether your deployment includes network egress controls and hardware monitoring that operate independent of model instructions. If those controls aren't in place, OpenShell and Sentry are one answer for what that layer can look like.
Routing agentic coding workloads after Sonnet 5.5's performance and safeguard changes
If Opus is your current tier for agentic coding tasks, Anthropic's Sonnet 5.5 release puts both routing reasons under review. Anthropic says Sonnet 5.5 outperforms Opus 5.5 on agentic coding benchmarks at lower per-task cost. Anthropic says the safeguard tier is now the same: Sonnet 5.5 receives the same cyber safeguards as Opus and Fable. The decision splits into two checks: whether Sonnet 5.5's benchmark holds for your specific workload, and whether safeguard parity changes the tier requirement for safety-driven routing decisions.
---