Agents Get Access. Infrastructure Gets Predictive. Robots Enter Production.

What happened
OpenAI launched GPT-5.5-Cyber in limited preview for defenders securing critical infrastructure and expanded Trusted Access for Cyber, a vetting framework that lowers refusals for approved defensive work like vulnerability triage, malware analysis, and patch validation while still blocking credential theft, stealth, persistence, and malware deployment. OpenAI also said some trusted users will need phishing-resistant account security starting June 1.

Why it matters
The shift here is not just capability; it is distribution. In effect, OpenAI is turning identity checks, workflow scoping, and permissioning into part of the product, which is likely how frontier labs will ship powerful dual-use agent systems without broad public release.

What’s next
OpenAI is positioning GPT-5.5 with Trusted Access for most defensive teams and reserving GPT-5.5-Cyber for narrower, higher-risk workflows like controlled validation and authorized red teaming. It is also tying the program to partners including Cisco, Cloudflare, CrowdStrike, Intel, Rapid7, Snyk, and others, so the next proof point is whether this translates into faster real-world detection and patching.

What happened
Kyndryl launched a Predict & Prevent capability on its Bridge platform that uses agentic AI to analyze telemetry across customers’ IT environments, detect patterns and surface potential failures before they cause outages. The feature delivers millions of actionable insights and can cut IT incidents by up to 50%, reducing root‑cause analysis time and saving customers millions.

Why it matters
Agentic AI is moving beyond chatbots to orchestrate complex infrastructure. By proactively detecting anomalies and recommending fixes, these agents demonstrate that autonomous workflows can improve resilience and lower costs for mission‑critical systems.

What’s next
Kyndryl plans to extend the capability across more industries and integrate it deeper into the Bridge platform. Expect other IT providers to build similar predictive agents as enterprises demand always‑on reliability.

What happened
Mozilla detailed how its Firefox team used Claude Mythos Preview and other models to harden the browser, saying the pipeline helped identify 271 bugs fixed in Firefox 150 and contributed to 423 security bug fixes shipped across April 2026. The company published a sample set that included long-dormant bugs, sandbox escapes, and other high-severity issues.

Why it matters
This is one of the clearest public proofs that agentic tooling is moving from “interesting demo” to production security workflow. Mozilla says the breakthrough came when models got better and agentic harnesses became able to create and run reproducible test cases, turning AI bug hunting from noisy speculation into usable signal.

What’s next
Mozilla says it plans to integrate this analysis into continuous integration so patches can be scanned as they land. If that happens, agentic security stops looking like a special project and starts looking like standard build-pipeline infrastructure.

What happened
Perplexity opened Personal Computer to all Mac users through its new Mac app after initially limiting it to Max subscribers on a waitlist. The tool lets agents work across local files, native Mac apps, the web, and more than 400 connectors, and it can also be run on an always-on Mac Mini and accessed remotely from an iPhone; the feature still requires a Pro or Max subscription.

Why it matters
This is a sign that OpenClaw-style personal agents are becoming a shipping product category, not just a developer experiment. Once agents can move across files, apps, web tools, and personal context, the competition shifts from raw reasoning to whether vendors can make those permissions useful without making them reckless.

What’s next
Perplexity says its older Mac app will be deprecated in the coming weeks, which shows the company is consolidating around an agent-first desktop client. The next battleground is likely to be safe approvals, enterprise controls, and whether managed-device environments will allow these local-context agents at all.

What happened
Cryptocurrency miner IREN announced a strategic partnership with NVIDIA to deploy up to 5 gigawatts of DSX‑aligned AI infrastructure and granted NVIDIA warrants to invest up to $2.1 billion. NVIDIA CEO Jensen Huang called AI factories “foundational infrastructure,” while IREN aims to deploy 1 GW of capacity by 2027, moving from bitcoin mining into large‑scale AI compute.

Why it matters
The deal underscores a global race to secure compute capacity for generative and agentic workloads. Building gigawatt‑scale data centers reshapes energy markets, supply chains and who controls the future AI stack.

What’s next
Regulatory approvals and financing terms will determine how fast the capacity comes online. Similar partnerships between hyperscalers and specialist data‑center operators are likely as demand for AI compute explodes.

What happened
Luxury conglomerate OTB Group partnered with Google Cloud to launch AI‑powered virtual try‑on experiences for Diesel and Jil Sander shoppers. Using Gemini multimodal models and Google’s Virtual Try‑On API, the app offers hyper‑personalized styling, realistic fit previews, AI‑assisted image editing and generative video to help clients visualize outfits before buying. OTB’s founder said generative AI is a strategic lever and the company plans to expand the service to more brands and markets.

Why it matters
Generative AI is moving into luxury retail, promising to boost sales and reduce returns by letting customers virtually “wear” products. This deployment shows how companies can combine proprietary data with frontier models to create bespoke experiences.

What’s next
OTB will roll the technology out across its portfolio and geographic footprint. Competitors are likely to adopt similar virtual try‑on tools, raising questions about intellectual‑property rights and data privacy.

What happened
Semafor debuted Semafor Intelligence, an AI‑assisted editorial product that uses embedding models to analyze transcripts from the publication’s flagship World Economy conference and distill them into nine key themes. The custom tool, built with OpenAI’s Codex agent and Voyage AI embeddings, vectorizes hundreds of sessions to identify topic clusters; journalists then review the themes, curate quotes and write the final copy.

Why it matters
Newsrooms are experimenting with generative and agentic AI to process massive datasets. By automating tedious summarization while retaining human editorial judgment, Semafor shows how AI can augment journalism without replacing it.

What’s next
Semafor plans to apply the tool to future events, and other media organizations may adopt similar AI summarization workflows. The success of these tools will depend on transparency and maintaining reader trust.

What happened
Spirit AI and Bosch China formed a strategic alliance to industrialize a “Universal Brain” for robots. The partnership aims to build a real‑world data loop within Bosch’s factories and logistics centers, combining Spirit AI’s embodied VLA models with Bosch’s mission‑critical sensors and actuators. The collaboration includes benchmarking the Spirit v1.5 model against global standards and promises to accelerate validation and mass production of general‑purpose robots.

Why it matters
Moving beyond lab prototypes, the deal fuses AI brains with industrial hardware, signalling that general‑purpose robots are nearing real‑world deployment. Harnessing real‑world data and hardware synergy could unlock adaptable robots for manufacturing and logistics.

What’s next
Spirit AI and Bosch will integrate the universal brain into Bosch’s factories, with potential expansion across the supply chain. As embodied AI scales, safety standards, workforce impacts and data‑sharing agreements will come into focus.

Agentic AI is crossing a threshold from experimentation to operational infrastructure. Whether securing critical systems, predicting outages, curating intelligence, or coordinating robots on factory floors, the real differentiator is no longer just model capability — it’s who controls access, workflows, permissions, and real-world execution safely at scale.

Build your own “predict, detect, and act” workflow. Use Perplexity Mac App as a personal agent layer across files and apps, experiment with Claude or OpenAI Codex for summarization and bug analysis, then map how agent permissions, approvals, and telemetry could flow through your own organization. The emerging stack isn’t just smarter models — it’s agents connected to infrastructure, identity, and real-world systems.