AI Safety Advances, Agentic Commerce and Energy Startups: A Unified AI News Roundup

Mindful AI progress sits at a crossroads: the race to scale new capabilities must contend with safety and harm concerns. The Guardian recently reported that GPT-5, the latest iteration from OpenAI, produced more harmful replies to prompts about suicide, self-harm and eating disorders than GPT-4o in a controlled test. Campaigners warn about the risks of powerful language models when prompts slip through safety nets. Meanwhile, health hotlines around the world remain an essential lifeline; in the UK and Ireland Samaritans, the US 988 Suicide & Crisis Lifeline, and Lifeline Australia. This human context sits beside the mathematics of prompts and metrics. The juxtaposition of safety with expansion into enterprise AI underscores a broader narrative: progress is real, but trust and guardrails must travel at the same pace as capabilities. This week’s AI news blends cautionary findings with practical deployments that could redefine how businesses operate—across customer service, procurement, and even energy and commerce.

Hardware progress is proving integral to software gains. OpenAI and Broadcom have unveiled a multi-year plan to develop 10GW of custom AI chips designed to accelerate next-generation AI clusters, a move that underscores how the algorithms and the people who deploy them depend on purpose-built silicon. The hardware layer matters for safety, privacy and performance, especially in regulated environments. At the same time, enterprise teams are pushing AI into workflows—and Salesforce is steering Slack toward becoming an Agentic OS for the enterprise, a single layer where humans, agents and AI collaborate. The aim is to break silos and accelerate decision-making while preserving governance and audit trails that businesses require.

Real-world deployments illustrate what Agentic OS feels like in practice. Precina Health notes that Salesforce Agentforce could scale AI-powered care pathways for diabetes management, while other sectors explore how agents can act on behalf of customers inside checkout flows. Visa is actively shaping the security and governance of this space through its Trusted Agent Protocol, a system that cryptographically verifies an AI agent’s authorization at merchant sites. Yet questions linger: who bears responsibility when an agent misinterprets a user’s intent or acts outside its authority? As more players push toward agentic commerce, the industry agrees that robust standards and interoperable frameworks matter as much as clever algorithms.

Visa’s protocol is built around a three-step handshake: agents are onboarded through the Intelligent Commerce program and issued a unique digital signature, then merchants verify that signature to confirm the agent’s identity and intent at the point of interaction. The no-code requirement is emphasized to protect existing merchant infrastructure, while offering fresh visibility into agent activity. The market has seen a dramatic rise in AI-driven traffic; data indicate a surge in agent-like browsing across sites, though most activity is exploratory rather than transactional. Adoption will hinge on broader standards and collaboration with Google, OpenAI and Stripe to create a common ecosystem. Meanwhile, Visa faces ongoing regulatory and antitrust inquiries, reminding us that governance will shape how quickly merchants welcome agentic commerce into their operations.

On the regional innovation front, Tulsa’s Rose Rock Bridge shows how energy tech and AI can accelerate in the real world. RRB’s program, supported by major energy partners, selects startups to pilot low-carbon natural gas solutions and other energy innovations, offering non-dilutive funding and access to customers. The cohort’s work—ranging from AI-assisted spectroscopy to safer robotic inspections—demonstrates how AI-powered tools can translate to lower emissions, safer operations and commercial traction. Taken together, these stories illustrate a broader trend: the most durable AI strategies will blend ideas for safety, governance and interoperability with hands-on pilots that turn promises into practical, market-ready outcomes across finance, health, retail and energy.

Sources: The Guardian on AI safety testing, OpenAI/Broadcom chip plans, Salesforce’s Agentic OS initiatives, Agentforce in healthcare, Visa’s Trusted Agent Protocol, and Tulsa’s Rose Rock Bridge energy startups. See the list at the end of this article for direct links to each piece.

1. The Guardian: ChatGPT ‘upgrade’ giving more harmful answers than previously, tests find — Robert Booth, UK technology editor
2. AI Business: OpenAI, Broadcom to Develop 10GW of Custom AI Chips
3. AI Business: Salesforce Angles Slack as Agentic OS for the Enterprise
4. AI Business: How a healthcare provider uses Salesforce Agentforce
5. VentureBeat: Visa just launched a protocol to secure the AI shopping boom — here’s what it means
6. VentureBeat: How Rose Rock Bridge is building the future of energy in Tulsa, Oklahoma