Anthropic logs queries on chikungunya transmissibility and H5N1 virulence in safety report

Executives at Anthropic and OpenAI publicly aligned on slowing frontier progress after researcher exits and statements citing >10% probability of AI-driven extinction this decade. The 2022 Urbina et al. Nature Machine Intelligence paper demonstrated a repurposed molecule generator producing 40,000 candidate chemical warfare agents in six hours, exceeding VX toxicity in multiple cases. Current DNA synthesis screening and model refusal layers remain bypassable through iterative prompting, as documented in the Anthropic red-team findings.

DIY synthetic biology kits and public LLM access lower barriers for individuals seeking pathogen engineering guidance. Existing blue-teaming exercises and voluntary API filters lack coverage against novel recombination strategies or non-English query variants. The 2024 IARPA-funded BENGAL benchmark showed frontier models retaining 68% of actionable steps on select-agent protocols even after safety fine-tuning.

Operational response requires mandatory sequence screening at synthesis foundries plus real-time model telemetry shared with national biosecurity agencies. Without enforceable thresholds on query persistence, current voluntary commitments will not scale against determined actors. Next regulatory filings from Anthropic and OpenAI must include refusal-rate metrics on high-risk pathogen categories.

Deployment of verifiable watermarking on generated sequences and cross-model query logging offers the clearest near-term control layer.