Today's news is dominated by a clear theme: AI is moving from individual chat assistant to embedded teammate. Anthropic put Claude inside Slack as a taggable coworker, Samsung handed ChatGPT and Codex to its global workforce, and OpenAI's security model began writing and shipping its own patches. Underneath the product news, the infrastructure arms race intensified, with multi-billion-dollar compute and memory deals and a notable Nvidia design that could nearly eliminate data-center water use. Below is a synthesis of the day's most significant developments.
Model and Product News
Anthropic introduced Claude Tag, a collaboration system that lets teams assign work to Claude directly inside Slack. Once added to selected channels and connected to a company's tools, databases, and codebases, the AI can be summoned with "@Claude" to complete tasks asynchronously. It retains channel context over time, learns from approved conversations, schedules long-running jobs, and follows up on unresolved work. Anthropic says its own teams now use the system to write 65% of their internal product code, and researcher Andrej Karpathy described the interaction model as a third major paradigm shift in software UI/UX. Alongside it, Anthropic is moving background task execution off local machines and into the cloud via Claude Cowork, and is consolidating access to its models across enterprise platforms.
OpenAI pushed automated security past detection and into remediation. Under a defensive initiative called Daybreak, the company released its full GPT-5.5-Cyber model to verified defenders along with an updated Codex Security plugin. Rather than simply flagging vulnerabilities, the system runs deep scans, builds threat models, reasons about how an attacker might exploit a flaw, and writes targeted patches that export directly into developer tooling through formats like SARIF and CodeQL. OpenAI cites an 85.6% score on the CyberGym benchmark, which it says outpaces Anthropic's Mythos 5 model. To deploy fixes responsibly, the company launched a Cyber Partner Program and a "Patch the Planet" effort with firms including Trail of Bits and HackerOne, and reports early contributions to projects such as Python, Go, cURL, and Sigstore.
OpenAI also rolled out Record and Replay for Codex, a feature that lets a user perform a task once on their computer while Codex watches, then converts the demonstration into a reusable skill. Notably, it captures intent rather than pixel coordinates by generating a natural-language instruction file the model reinterprets against whatever the interface looks like at run time — an approach meant to avoid the brittleness of traditional robotic process automation tools.
In media generation and research models, Seedance 2.5 can now produce 30-second native 4K video from a single prompt, Mistral launched OCR 4 to structure messy document data for enterprise agents, and Baidu open-sourced an "Unlimited OCR" system aimed at long-document processing. Sakana AI released Fugu, which coordinates multiple large language models through a single API to act as one "super-agent," while a compact model called VibeThinker-3B reportedly matches larger systems with a fraction of the parameters.
Enterprise Adoption
Samsung Electronics is rolling out ChatGPT Enterprise and Codex to employees in South Korea and to all Device eXperience staff worldwide, in what OpenAI called one of its largest enterprise deployments to date. Samsung plans to apply the tools across R&D, software and product development, marketing, manufacturing, and corporate functions, and emphasized use by non-technical teams to turn ideas into internal tools, websites, and automated workflows. The breadth is the story: rather than equipping a small group of developers, Samsung is positioning AI as a default layer across roles.
Partnerships extended into infrastructure operations as well. Google Cloud and Nokia expanded their alliance with Gemini-based AI agents for autonomous telecom network operations, and Microsoft outlined a vision of "agentic observability" in which AI agents reason over telemetry and incidents to detect and resolve cloud issues faster than dashboards and alerts. OpenAI, meanwhile, struck a licensing deal with Getty Images to surface licensed commercial and editorial photography inside ChatGPT search results — improving image quality while reducing reliance on contested web scraping.
Infrastructure and Compute
The competition to lock down compute and memory escalated. Reflection AI agreed to buy access to Nvidia GB300 infrastructure from SpaceX's Colossus 2 data center in a deal worth up to $6.3 billion, underscoring how AI labs are increasingly securing long-term capacity outside the traditional hyperscalers. Micron and Anthropic signed a multi-year strategic agreement covering memory and storage for Anthropic's Claude data centers and joint work on AI-specific memory architecture; Micron will also use Claude internally and is investing in Anthropic's Series H round. Demand for high-bandwidth memory has been strong enough that Micron joined the trillion-dollar valuation club as labs move to secure supply.
On sustainability, Nvidia said its upcoming Rubin data-center reference design replaces conventional cooling with a fully liquid-cooled architecture running at higher temperatures, allowing outdoor dry coolers to reject heat for much of the year. The company claims this cuts water consumption from roughly 2.6 million gallons per megawatt per year for cooling-tower systems to near zero, though it did not disclose the additional construction costs.
Security and Governance
A high-severity SSRF flaw in Cisco Unified CM (CVE-2026-20230) is now being actively exploited, with attackers attempting to write files to vulnerable systems; Cisco patched it earlier this month, but unaddressed deployments risk root-level compromise. Separately, a survey found that organizations using AI more heavily also report more AI-related security incidents — a correlation that reinforces the case for access controls, monitoring, and governance before AI tools spread across workflows.
Quick Takes
Google added an optional AI health coach to its fitness tracker; the coach can be ignored entirely, but getting useful guidance requires extensive manual input of medical history and test results.
Engram raised $98 million from Sequoia to build AI with permanent memory.
Perplexity integrated China's flagship GLM-5.2 model into its new agent ecosystem.
Cursor quietly acquired Continue, an open-source alternative to GitHub Copilot.
An AI-powered law firm reportedly won its first court trial.
OpenAI is targeting a roughly $100 billion valuation and made its debut in the advertising market at Cannes.
Cerebras's CEO confirmed OpenAI's GPT-5.4 runs on Cerebras hardware, with a GPT-5.5 rollout planned.
Baseten raised $1.5 billion to scale lower-cost alternatives for serving large models.
Oklo partnered with Centrus to fuel Sam Altman-backed nuclear-powered data centers.
Running tallies of 2026 tech layoffs increasingly cite AI as an employer-stated factor.
What This Means for Your Business
The most actionable shift today is the arrival of AI as a persistent team member rather than a personal tool. Claude Tag and Samsung's broad ChatGPT and Codex rollout point to the same operating model: AI that lives where work already happens — Slack, internal tooling, existing codebases — and is accessible to non-technical staff. For a small or mid-sized business, the practical takeaway is to stop thinking about "who gets a license" and start thinking about which recurring workflows could be handed off to an assistant that retains context. The barrier to entry is no longer technical skill; it is identifying the repetitive, well-documented tasks worth delegating.
OpenAI's Record and Replay feature lowers that barrier further. The ability to demonstrate a task once and have it become a reusable, intent-based skill is well suited to the messy, undocumented processes that dominate small-business operations — onboarding steps, report assembly, data entry across systems that lack clean integrations. Because the approach interprets intent rather than fixed screen positions, it should prove more durable than older automation that breaks whenever a vendor updates an interface. Teams should begin cataloging the manual processes that consume the most time, as these are now candidates for low-effort automation.
The security news cuts both ways and deserves sober attention. OpenAI's patch-writing system and its "Patch the Planet" program suggest that defensive tooling is becoming dramatically more capable, which benefits resource-constrained organizations that cannot staff a security team. At the same time, the actively exploited Cisco flaw and the finding that heavier AI adoption correlates with more incidents are reminders that expanding AI access widens the attack surface. The right posture is to pair adoption with governance from the start: clear access controls over what AI agents can read and act on, logging of their actions, and human review of anything that touches production systems or customer data.
Finally, the infrastructure and cost signals matter even for businesses that never touch a data center. Multi-billion-dollar compute and memory commitments, new lower-cost serving providers like Baseten, and OpenAI's move into advertising all point to an industry still searching for sustainable economics. For buyers, that argues for flexibility — avoid deep lock-in to a single model or vendor, watch for pricing changes as providers test new revenue models, and treat today's capabilities as a moving baseline rather than a finished product. The companies that benefit most will be those that build repeatable internal processes around AI while keeping the option to switch the engine underneath.