/
← Voltar ao feed
low

Cloudflare previews Clairic, a System One extraction model on Workers AI (no security incident)

1 fonte(s) · 1 atualização(ões) · publicado 19/10/2026 14:48 · coletado há 1 dia

Cloudflare announced a preview of Clairic, the first "System One" extraction model, hosted on Workers AI as @cf/cloudflare/clairic, with open weights planned. The model extracts, computes, estimates and forecasts typed values from unstructured text and images in a single forward pass. It generates no output tokens and returns confidence scores, p_null and 80% intervals. It builds on the System One concept introduced by TypeSafe (the Jev model) and complements Cloudflare's decision model Clef. This is product news with no vulnerability, threat actor or exploitation described, so it has no direct threat impact.

Technical Details

  • Clairic accepts text plus up to four images and a map of typed fields (text, number, date, duration, list, estimate, forecast) and answers all fields in parallel passes.
  • Cloudflare claims it cannot produce malformed output, though it can still be wrong, and returns null when a value is absent.
  • In Cloudflare's benchmark, a Clairic + Clef + LLM cascade ran 3.2x faster and cost 49% less than an all-LLM pipeline using GLM-5.3.
  • The article argues that feeding an LLM typed values rather than raw email reduces exposure to prompt injection.
  • Stated limitations: multi-step arithmetic is weak, estimates need evidence, intervals can run narrow on unfamiliar data, and forecasting is basic.

Remediation

No remediation applies. Defenders adopting the model should rely on confidence, p_null and interval outputs to decide what to automate and what to send for human review.

Linha do tempo

  1. blog.cloudflare.com ↗ · pub. 19/10/2026 14:48 · coletado 19/10 18:00 primeiro registro

    Cloudflare previews Clairic, a System One extraction model on Workers AI (no security incident)

    ler a análise desta fonte

    Cloudflare announced a preview of Clairic, the first "System One" extraction model, hosted on Workers AI as @cf/cloudflare/clairic, with open weights planned. The model extracts, computes, estimates and forecasts typed values from unstructured text and images in a single forward pass. It generates no output tokens and returns confidence scores, p_null and 80% intervals. It builds on the System One concept introduced by TypeSafe (the Jev model) and complements Cloudflare's decision model Clef. This is product news with no vulnerability, threat actor or exploitation described, so it has no direct threat impact.

    Technical Details

    • Clairic accepts text plus up to four images and a map of typed fields (text, number, date, duration, list, estimate, forecast) and answers all fields in parallel passes.
    • Cloudflare claims it cannot produce malformed output, though it can still be wrong, and returns null when a value is absent.
    • In Cloudflare's benchmark, a Clairic + Clef + LLM cascade ran 3.2x faster and cost 49% less than an all-LLM pipeline using GLM-5.3.
    • The article argues that feeding an LLM typed values rather than raw email reduces exposure to prompt injection.
    • Stated limitations: multi-step arithmetic is weak, estimates need evidence, intervals can run narrow on unfamiliar data, and forecasting is basic.

    Remediation

    No remediation applies. Defenders adopting the model should rely on confidence, p_null and interval outputs to decide what to automate and what to send for human review.