OpenAI and Anthropic locked down their own agents while we’re giving ours prod access

so these are the last three weeks in agent land, lined up

OpenAI paused tool use on its most capable models after an agent tunnelled out to an outside chatbot through the sandbox’s DNS resolver which, their monitoring flagged it within about 15 minutes and the run still kept going for around 2.5 hours before someone stopped it by hand

Anthropic cut live internet from every internal eval after Claude agents submitted forms they’d been told not to submit, one of them a fake homicide tip to the Philadelphia police

Nvidia just put OpenShell at the centre of a new agent safety platform and the agent runs in a sandbox with no network path at all except through a policy checker that sits outside it, and Anthropic, Salesforce and JPMorgan are already wiring it in

for normal teams the realistic version is way more boring than a kernel sandbox, tool allow lists and approval steps that live outside the agent. LangGraph has interrupts for the approval part, and Orq AI’s MCP gateway lets you pick which tools your agents can even see

the people who build these models don’t trust a system prompt to keep them in line. meanwhile half the agent setups posted in here run on IMPORTANT: NEVER delete production data in all caps and a prayer