Anthropic

Anthropic found evaluation runs that reached real systems after models were told they were in a simulation. A prompt described isolation that the network did not enforce.

Anthropic, Macquarie Asset Management and GIC have created a vehicle meant to finance, build, operate and lease AI data centres, initially in the United States. The structure can turn a long-term computing commitment into an investable asset, yet the announcement omits the first facts that would turn the platform into infrastructure: a site, committed project capital, grid terms, a construction date and energised load.

Bloomberg identifies Anthropic as the AI lab behind Volta’s announced $10bn strategic partnership. Volta and Bitdeer do not name the lab, so that identity remains reported rather than first-party confirmed. What the first parties do disclose is unusually detailed: a 121MW IT lease at Bitdeer’s Tydal campus, about $4.7bn of scheduled payments over 16 years, a ten-year termination right, conditional credit support and two delivery dates. The important story is not one giant price; it is the chain of conditions that must make the compute usable.

Anthropic marked a minor Claude Sonnet 5 performance incident resolved at 07:04:49.691 UTC on 31 July, 46 minutes and 41.623 seconds after it opened. That short chronology establishes a start and an end, but almost nothing about the customer experience between them: no symptom, access surface, geography, request count or technical cause was published.

A third-party test misconfiguration gave Claude internet access, allowing three models to reach live systems while completing simulated hacking exercises.

The notable number in the reported financing talks for a data-centre campus in Hubbard, Texas, is not only the $15bn banks may lend. It is the obligation that could sit behind the debt. The Wall Street Journal reported that a Morgan Stanley-led consortium was discussing financing for Nexus Data Centers, while Google would guarantee billions of dollars of Anthropic’s lease and power commitments if the startup defaulted. Nothing in the public record shows that the financing has closed. Yet the proposed structure already reveals how AI infrastructure is moving beyond ordinary cloud procurement: lenders may be underwriting a chain that runs from a privately held model developer, through a project company and an onsite power plant, to the credit of a much larger technology group.

Anthropic says a retrospective review of 141,006 cybersecurity-evaluation runs found three incidents, spanning six runs, in which Claude models reached the open internet and compromised systems belonging to three real organisations. The models had been told they were inside sealed simulations. The network did not enforce that claim. This is not evidence that Claude formed an independent objective or deliberately escaped: the disclosed models continued capture-the-flag tasks under a false account of their environment. It is evidence that a safety boundary expressed in a prompt, and assumed by two organisations, is not a boundary until infrastructure makes it true.

Anthropic marked a cross-model Claude incident critical after four product surfaces moved into major-outage status. Its final timeline bounds elevated errors to 101 minutes, yet the record never says how many requests or customers failed. That gap turns a familiar status page into a sharper question about what providers must disclose when one model service becomes infrastructure for other businesses.

Anthropic released computer use in public beta on 22 October 2024 with an upgraded Claude 3.5 Sonnet. Developers could let the model inspect screenshots and propose mouse and keyboard actions through the Anthropic API, Amazon Bedrock and Google Cloud Vertex AI. The launch was deliberately experimental: Anthropic called the capability cumbersome and error-prone, reported only 14.9% on OSWorld's screenshot-only tasks, and told developers to begin with low-risk work.

Anthropic open-sourced the Model Context Protocol on 25 November 2024 as a common way for AI applications to connect to external data and tools. The launch included a specification, SDKs, local server support in Claude Desktop and reference servers. MCP could reduce repeated connector work by defining shared messages and capabilities, but the launch specification explicitly said the protocol could not enforce consent, authorization or data protection on its own. Standardized connectivity made the trust boundary easier to describe; it did not make every server safe or every tool call correct.

Anthropic released Claude 3.7 Sonnet on 24 February 2025 with two modes in one model: a standard response and an extended-thinking mode. API users could set a token budget for the longer analysis, making latency and cost part of the task configuration rather than a reason to choose a separate model. Anthropic reported stronger coding and agent results, but its published scores depended on tools, test subsets and high-compute scaffolds. The launch established a controllable reasoning budget, not proof that longer thinking was always more accurate or trustworthy.

Anthropic's first Economic Index, announced on 10 February 2025, mapped a short snapshot of Claude.ai conversations to occupational tasks in the US O*NET database. Its best-known figures—36% task coverage across occupations and a 57% to 43% augmentation-automation split—describe patterns inside that platform sample. They do not count workers, jobs, hours, productivity gains or employment losses.

Anthropic resolved elevated errors on Opus 4.8 and Haiku 4.5 at 15:23:37 UTC, then opened a Sonnet 5 error record at 15:24:05. Both minor incidents named the same five service surfaces. The 28-second gap creates operational continuity for customers, but it does not prove one continuous outage or a shared cause.

Haiku 4.5 returned from monitoring to an identified problem, a critical multi-model error interval overlapped it, Claude workflow features then failed under a separate critical record, and Opus 4.1 recorded another incident the next morning. The sequence shows concentrated operational risk without proving a common technical failure.

Reported talks over a two-year compute lease would give Meta a first large external customer and Anthropic another source of capacity, but the proposed $10 billion ceiling is not a contract, committed revenue or proof that Meta has spare infrastructure.