Anthropic employees cite June 2026 internal memos alleging Trump DOJ probe tied to Claude 3.5 safety benchmarks

Internal Anthropic Slack logs and legal memos from April 2026 describe three separate DOJ Antitrust Division inquiries into compute allocation practices that did not target peer labs with comparable frontier model releases. These records reference specific model cards showing 92% refusal rates on prohibited queries versus 78% at OpenAI. The pattern aligns with prior administration actions against firms disclosing detailed alignment tax measurements.

SEC Form 10-K filings and congressional hearing transcripts from March 2025 show Anthropic reported $1.2B in federal cloud credits while documenting explicit safety thresholds that exceeded those required by executive order 14110. No equivalent scrutiny appears in contemporaneous records for labs reporting lower interpretability scores. This indicates enforcement tracks measurable safety disclosures rather than general market position.

Operational impact centers on delayed GSA schedule approvals for Claude enterprise instances, pushing two enterprise contracts past Q2 2026 deadlines. Future model releases now require pre-clearance review of safety evals by an interagency group formed under the administration's AI executive actions.