Anthropic says the two products use the same underlying model with different safety controls: Fable 5.1 is generally available, while Mythos 5.1 is limited to trusted-access programs for cybersecurity and life-sciences work. Anthropic Anthropic estimates that Fable 5.1 costs 25% less than Fable 5 for typical workloads priced by the amount of text processed, with savings up to approximately 45% for work where AI completes many steps on its own. Its enterprise data-protection system stores customer data in cloud infrastructure controlled entirely by the customer and is scheduled to reach enterprise customers in phases beginning later this fall; eligible customers can use a policy that keeps no customer data after processing until then. Anthropic Anthropic also says its newest cybersecurity safety controls block 60% fewer mistaken warnings than before and allow finding software weaknesses without building attacks from those weaknesses. Anthropic CEO: Business strategy & enterprise transformation: Tie any rollout of an advanced AI model to a business workflow, a decision about who controls the data, and a measurable outcome. Approve the higher-capability version only where the company can fund the controls that make its use defensible. CTO / CIO: Technical posture: Test the customer-controlled data path, retention settings, and fallback model before production. Record which safety controls are active in each environment so a model upgrade cannot silently change the risk profile.

OpenAI defines that threshold as the ability to identify and develop functional attacks that use previously unknown software weaknesses across many well-protected real-world systems that matter to essential services without human intervention, or to devise and execute end-to-end novel attacks against hardened targets from a high-level goal. OpenAI OpenAI reported that Astra scored 100% on its ExploitBench benchmark for developing exploits from known vulnerabilities. On an internal benchmark containing 20 high-severity vulnerabilities disclosed more recently, OpenAI said Astra found and used two zero-day vulnerabilities in exploit chains. Testing led by security experts also found a sequence of steps that took over a browser that escaped a sandbox and a sequence of steps that gained higher access on a computer from a regular user to full computer control. OpenAI OpenAI said those results reflect Daybreak Blue access rather than the default production configuration and described additional refusal training, controls against harmful use, and monitoring that can stop potentially unauthorized activity. OpenAI CEO: Business strategy & enterprise transformation: Put cyber-capable AI on the enterprise risk map before it enters a workflow. Fund defensive use cases and independent testing together, with a named executive accountable for misuse and containment. Board: Governance & accountability: Require management to report the highest capability tier reachable through company accounts, the controls that separate defensive from offensive use, and the incident trigger that forces human intervention.

The program includes Red Tempest, a model that simulates an attacker for advanced attack scenarios, and Blue Solano, a model that helps protect enterprise systems. CrowdStrike CrowdStrike says SafeMind is trained on automatic security data from Falcon sensors, threat intelligence, Falcon Complete MDR event annotations, and 15 years of hands-on work responding to security incidents. It uses NVIDIA Nemotron open models with NVIDIA as AI design partner and CoreWeave for training and inference. Its software that lets AI tools work together pit the red and blue models against each other in a closed loop, while also supporting advanced and openly available AI models. CrowdStrike CrowdStrike reports 29% higher detection, six-times-faster security fixes, and 99% cost savings on detection and remediation in its evaluations. CrowdStrike CrowdStrike also launched an AI Partner Specialization that gives partners paths to build, resell, manage, and deliver AI-powered security, plus a Verified Agent certification for AI systems built by partners. CrowdStrike CEO: Business strategy & enterprise transformation: Identify the operating loop where proprietary data and a closed feedback cycle can create a durable advantage. Set the outcome metric before choosing whether to build, buy, or partner for the model. Market transformation: Industry-level shift: Security vendors are becoming distribution and certification layers for partner-built AI. Map which partners can reach your customers, which platform owns the trust signal, and where switching costs will accumulate. COO / Chief Transformation Officer: Require a human-approved action boundary for remediation. Measure detection quality, time to contain, mistaken warnings, and business downtime, not the number of alerts an AI system touches.

The team reported clean evaluation scores and passing unit tests. VentureBeat When a low-access account asked the same questions that a high-access account had already asked, the assistant returned SharePoint content the lower-access user could not open directly. VentureBeat reported that the custom system that finds information for an AI answer bypassed Azure AI Search's native checking each user's permissions when information is retrieved. VentureBeat Microsoft documents that access rules for individual documents is disabled when the permitted-groups field is not mapped, which makes checking whether a person is allowed to see the information an explicit design requirement rather than a default assumption. Microsoft Learn CEO: Business strategy & enterprise transformation: Make data access and business outcome co-equal launch criteria. No AI workflow that reads company information should move to production without identity-specific tests and a named data owner. CTO / CIO: Technical posture: Use permission checks when information is retrieved or build an equivalent check into every custom pipeline. Test low-, medium-, and high-access identities against identical prompts, preserve the information-lookup logs, and fail closed when permission metadata is missing. Board: Governance & accountability: Treat an AI information-access leak as an permission incident, not a model-quality defect. Require quarterly evidence that high-impact AI workflows enforce the same permissions as the systems they query.

Hosted on Microsoft Azure, the companies are initially targeting one million users across the Middle East and Africa. PR Newswire HUMAIN ONE is described as an AI layer that lets employees interact with business software, company information, and work processes through a unified interface. The companies plan to extend the offering into HR, finance, procurement, and IT modules. PR Newswire HUMAIN plans to make its first AI PC available for enterprise purchase beginning September 20, 2026, with an initial target of one million AI PCs across the Middle East and Africa by 2030; the device is designed to run AI workloads locally alongside cloud services. PR Newswire CEO: Business strategy & enterprise transformation: Evaluate regional AI bundles as distribution and control decisions, not feature purchases. Negotiate portability, data-use limits, and exit rights before the first workflow is scaled. CMO / Chief Strategy Officer: Anchor the market message in a measurable customer result and local trust requirement. Use language, regulation, and service context as differentiation only where customers will pay for the difference. Market transformation: Industry-level shift: Bundles are pulling models, workplace software, devices, and regional services into one buying motion. Map which layer owns the customer relationship and which layer captures the usage data that improves the bundle.

Reuters said the arrangement covers about 350 megawatts of capacity for a Texas data-center project developed by Hut 8, with NVIDIA capacity intended to meet demand for Claude. Reuters The reported deal is a six-year commitment, according to coverage of the source-based agreement. Reuters It turns model demand into a physical delivery problem involving power, buildings, chips, network capacity, and schedule risk. CEO: Business strategy & enterprise transformation: Build a model-access portfolio before growth makes one provider unavoidable. Assign each critical workflow a performance tier, a fallback path, and a maximum acceptable migration window. CFO: Capital allocation & economics: Price the full service, not the per-text processing fee. Include reserved capacity, migration work, latency, energy exposure, and outage impact in every provider comparison. CTO / CIO: Technical posture: Keep the execution layer replaceable through automatically choosing which AI model handles each task, portable prompts, observability, and tested failover. Track where a contract for computer capacity creates a technical dependency that procurement alone cannot unwind.

The program gives partners routes to build, resell, manage, and deliver AI-powered security with CrowdStrike, while helping customers move from AI adoption to production and scale. CrowdStrike CrowdStrike says the specialization spans infrastructure, models, data, applications, identity, development platforms, and services. Its Verified Agent certification validates AI systems built by partners against CrowdStrike requirements and provides a route to market through the CrowdStrike Marketplace. CrowdStrike CEO: Business strategy & enterprise transformation: Decide which AI features should be sourced through a platform ecosystem and which must remain company-owned. Make certification, portability, and incident accountability explicit in every partner agreement. Market transformation: Industry-level shift: The new market layer is a governed exchange between platforms, partners, and AI systems. Map where validation creates real safety evidence and where it is only a channel-control mechanism. CMO / Chief Strategy Officer: Sell trust as an evidence-backed service outcome. Give customers a clear explanation of what certification covers, what it does not cover, and who owns the result after deployment.