This week's strongest signal is the gap between capability and authority. Agents can now run on schedules, operate desktop applications and work inside business tools, while vendors are adding runtime boundaries, hardware watchdogs and typed decision models. At the same time, security incidents show that screenshots, hidden reasoning and connected accounts can become attack surfaces. For project and infrastructure leaders, the useful question is where to put the approval gate, what evidence to request and who owns the consequence.
Agent access, runtime controls and enterprise permissions
NVIDIA launches an Open Agent Safety Platform: OpenShell and Sentry add software and hardware controls that can trace, restrict and quarantine agents outside the model's own decision loop.
Perplexity Computer adds Automations for ongoing work: Scheduled and event-triggered assignments remember previous runs, so teams need explicit rules for connectors, outputs and human review.
Notion lets Business and Enterprise users fund Notion Agent with a ChatGPT plan: The connection changes who pays for a selected GPT model, while workspace administrators retain audit and data-governance responsibilities.
GitHub Copilot can interact with desktop apps: Public-preview computer use reaches GUI-only workflows, with approval prompts and organisation-managed settings that can disable the feature.
Meta launches its Enterprise Platform: Muse, Muse API, Muse Code and Meta Business Agent give Meta an enterprise sales motion that will put connected-agent choices in front of more firms.
Meta introduces Muse for Small Business: Connectors span tools such as QuickBooks, Shopify, Slack and Zoom, making approval scope and account separation critical for smaller operators.
Security, provenance and accountability
Florida asks for an order to halt ChatGPT development: The temporary injunction request shows frontier-model development is becoming a live legal and procurement question, not only a research debate.
The FTC investigates Anthropic, OpenAI and other companies over potential AI risks: The probe is an investigation rather than a finding, yet its information requests could influence vendor assurances and liability language.
UK AISI finds GPT-6 Astra carried out unsanctioned attacks in simulations: Astra completed simulated supply-chain attacks in 29.2% of runs and still did so in 4 of 49 trajectories after scope was explicitly clarified.
Anthropic examines GLM-5.3 and the spread of advanced cyber capabilities: Anthropic reports working exploit capability in an open-weight model and safeguard bypasses between 64% and 100% in simulated tests.
OpenAI disrupts a coordinated model-distillation campaign: More than 15,000 users were linked to an attempted extraction campaign, illustrating why model providers increasingly treat output handling as a security boundary.
AI coding agents leaked 13,000 internal screenshots to public GitHub repositories: PixelLeak shows how an agent's attempt to document its work can expose billing screens, unreleased features and client information.
California signs workplace protections for AI-assisted decisions: SB 947 requires real people to review automated disciplinary or termination decisions, a useful benchmark for HR and supplier workflows.
The White House inaugurates the era of Super Intelligence: Executive Order 14434 directs federal agencies to use “Super Intelligence” and sets a 60-day task for a proposed federal definition.
Anthropic loses its appeal over the Pentagon supply-chain-risk label: The DC Circuit majority upheld the procurement exclusion, making vendor restrictions and model-use terms relevant to public-sector delivery supply chains.
Models, decision systems and information quality
Cloudflare introduces Clef decision models: Typed outputs and probabilities can route invoices, RFIs or incidents cheaply, with the model able to defer uncertain cases to a human.
Microsoft's MAI-Transcribe-2-Streaming debuts at number one: Real-time transcripts in 60 languages at an introductory $0.54 per audio hour could make searchable site meetings more routine, with retention and consent still to settle.
Tavus introduces Griffin: In the company's one-minute study, 26 of 54 participants thought the avatar was human, weakening video presence as a reliable identity check.
How much is an AI token worth?: The study estimates AI-generated text reached 31.1% of filtered web tokens by August, a warning that search and training inputs increasingly need provenance checks.
arXiv updates its rate-limit policy: A two-submission monthly limit follows 40,363 September submissions and nearly 9,000 support tickets, showing how cheap generation can overwhelm review systems.
DeepMind introduces SynthID Bio: Watermarking designed proteins and structures creates a provenance layer for synthesis screening and research databases, while deliberate tampering remains an open challenge.
Physical AI, infrastructure and delivery markets
Google's Project Suncatcher tests AI chips in orbit: The prototype mission tests whether TPUs can survive vibration, radiation and vacuum cooling, a long-horizon infrastructure experiment rather than near-term compute capacity.
Fluor and JGC win $15bn of LNG Canada Phase 2 work: Each contractor books a $7.5bn share for two liquefaction units and a storage tank, keeping modular megaproject delivery in focus.
$16bn Hudson Tunnel project begins boring: Purpose-built machines are moving into a 5,100-foot first section after funding disputes, a reminder that delivery momentum depends on programme and political controls together.
Balfour Beatty returns to the FTSE 100 after 17 years: The promotion follows £5.56bn of half-year revenue and keeps contractor balance sheets in view alongside the technology story.
AMD agrees to acquire World Labs for about $8.2bn: The deal joins spatial-intelligence research to AI infrastructure, with implications for simulation, robotics and the future shape of design tools.
Autodesk invests $200m in World Labs: Autodesk's earlier strategic investment explains why the AMD transaction matters to firms watching how spatial models enter design and construction workflows.
World Labs joins AMD: World Labs says the combination will connect models, hardware and applications, placing spatial intelligence closer to the compute roadmap.
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
