Anthropic Launches Claude Corps — FLI Safety Index Gives Anthropic C+, xAI a Failing Grade, All Labs Accused of Moving the Goalposts

Claude Corps — What It Is and Who Can Apply

Anthropic has announced Claude Corps, a paid 12-month fellowship designed to train future AI professionals within nonprofit organisations. The program is open to applicants aged 18 or older, with less than two years of full-time work experience, who are authorised to work in the United States. Fellows are placed inside nonprofits to build AI capability — implementing Claude tools, training staff, and developing AI-powered workflows for organisations that would otherwise lack the resources to bring AI expertise in-house.

The programme is Anthropic's clearest signal yet that it sees nonprofit and social sector AI adoption as a strategic priority — not just corporate and enterprise. The positioning makes sense commercially: nonprofits that build AI workflows on Claude during a fellowship year become long-term institutional customers. It also makes sense strategically: Anthropic has consistently framed its mission around beneficial AI deployment, and a fellowship programme that builds AI capacity in civil society organisations is a more credible expression of that framing than corporate partnerships alone. Fellows receive a salary, benefits, and access to Anthropic's tools and training resources for the full 12-month placement.

FLI AI Safety Index — What the Grades Mean

Anthropic ranked first in the Future of Life Institute's latest AI Safety Index, but received only a C+ overall. OpenAI and Google DeepMind each received a C. Meta improved to fourth from sixth, while xAI fell to seventh from fourth. xAI, DeepSeek, and Mistral received failing overall grades. The FLI index assesses labs across dimensions including safety research investment, governance commitments, red-teaming practices, incident reporting, and compliance with prior commitments. Anthropic's lead reflects its Responsible Scaling Policy and its comparative willingness to publish safety research — but the C+ grade signals that even the highest-rated lab falls significantly short of what the FLI considers adequate.

The reviewers said Anthropic, OpenAI, Google DeepMind, and Meta have weakened or eliminated earlier commitments to pause development if their systems approached specified danger thresholds. The panel described the changes as "moving the goalposts" and said they have undermined safety frameworks across the industry. FLI chair Max Tegmark said: "AI companies are sprinting toward a cliff. Despite acknowledging the great risks of artificial superintelligence, they continue racing to build it."

xAI's Fall From 4th to 7th — Why It Matters

xAI's drop from 4th to 7th with a failing grade is the most significant movement in this year's index. The FLI's assessment covers: safety research publication, governance commitments, incident reporting, red-teaming, and compliance with prior commitments. xAI has published no safety research comparable to Anthropic's Constitutional AI or Interpretability work. It has not published a system card for Grok 4.5. It has no documented red-teaming framework equivalent to OpenAI's Preparedness Framework or Anthropic's Responsible Scaling Policy. The adult content controversy — over 50% of Grok traffic confirmed as adult content — and the $530 million legal reserve disclosed in the SpaceX IPO prospectus are reputational factors that the index separately documents.

The practical implication for enterprise AI buyers: FLI grades are not procurement criteria, but they are a useful proxy for governance maturity. A failing FLI grade from a lab that also has no system card for its latest model (Grok 4.5), no published red-teaming framework, and a $530 million legal reserve for AI-related litigation is a combination that enterprise compliance teams will need to document when evaluating Grok as a production API provider.

The "Moving the Goalposts" Accusation — What Changed

The FLI's core accusation is specific: Anthropic, OpenAI, Google DeepMind, and Meta all previously made commitments to pause development if their systems approached specified danger thresholds. Those commitments have been weakened or eliminated. The pattern: each lab made commitments in 2023-2024 when capability thresholds felt distant. As models approached those thresholds in 2025-2026, the commitments were quietly revised upward or reframed. JADEPUFFER — the first autonomous AI ransomware attack documented just days ago — is a concrete example of a capability threshold being crossed in the wild while the labs that trained the enabling models were modifying their safety frameworks.

Sources: Medium (AI Week Jul 6-12), Mark McNeilly Substack July 10 edition · Related: JADEPUFFER: first autonomous AI ransomware → · Anthropic's government commitments on AI safety → · Grok 4.5 review →