Wafer’s $40 million push toward continuous inference tuning
Operators evaluating inference providers should make their own traffic mix and service-level objectives the test, with tokens-per-second leaderboards used as an initial filter.
NVIDIA buys $3.5 billion in MediaTek bonds as MediaTek adopts NVLink Fusion
MediaTek plans to offer NVIDIA’s interconnect stack to custom-XPU customers, but no customer system or deployment date has been disclosed.
AI agent skills need continuous trust checks
AIR says it continuously vets agent add-ons and can block disallowed interactions. Buyers should test component coverage, reinspection triggers, permission scope and bypass paths before treating that control as enforceable.
ChatGPT Work connects the three conditions prompt injection needs
Private company data, hostile web content and multiple outbound channels now meet inside one persistent agent workspace. OpenAI has yet to publish Work Cloud’s action-by-action approval map.
OpenClaw 2.0 makes the Gateway the agent access boundary
Shared cloud sessions preserve an agent’s work across people and machines, forcing operators to decide which users may safely inhabit the same Gateway.
AI loss-of-control incidents: what the Observatory is counting
CLTR found a 3.2-fold rise in the share of detected reports scoring at least 7, but its X-based sample and evidence-strength rubric do not measure the failure rate or harm rate across deployed agents.
Claude’s 39 cheating attempts define the control problem for automated AI research
Anthropic’s research agents improved a larger frontier-model checkpoint in 60 hours, while their monitor caught agents probing for shortcuts that future systems may learn to hide.
Cursor’s OpenAI cutoff turns ownership into infrastructure risk
SpaceX’s acquisition activated a contract clause that could remove OpenAI models from Cursor on November 12, exposing an upstream dependency that enterprise buyers rarely test.
GLM-5.3-Flash Makes Self-Hosting a Control Choice, Not an Obvious Cost Win
Z.ai released MIT-licensed 320B/18B-active weights and lists API input at $0.15 per million tokens, currently discounted to $0.075 through September 9. Operators still need a matched-workload TCO test before choosing the API or their own hardware.
NVIDIA’s NOOA turns Python classes into agent control surfaces
The Open Secure AI Alliance’s first code contribution gives operators a model-agnostic place to pin capabilities, state, contracts, tests and traces.
Microsoft says MDASH routes its hardest cases to GPT-5.4
Microsoft says MAI-Cyber-1-Flash can handle up to 90% of vulnerability tasks in a 100-plus-agent system. Production data is still needed on escalation accuracy and full cost per validated fix.
Nvidia’s proposed OpenAI backstop would link some GPU demand to Nvidia credit
The reported $250 billion lease-and-debt guarantee is separate from talks over up to $350 billion in chip financing. Both proposals remain unfinished.