GovBench

Aligning AI to government

[GovBench @GovBench]

GovBench issued 2 urgent advisories:
1️⃣ Block China-hosted LLM services on government networks: https://t.co/Vqpjdoe9iV
2️⃣ Ban China-trained LLMs from production government systems: https://t.co/QKHmwAKFYb

[GovBench @GovBench]

Now that we have concluded briefing senior US National Security leadership, we're revealing our findings.

🚨 GovBench tested Chinese vs. Western LLMs on U.S. Military domains. We issued 2 National Security Advisories.

When we added Chinese-censored words into the prompts, the results were shocking. 👇

[GovBench @GovBench]

This shows hidden mode-switch behavior:
🔒 Refusals
📄 Boilerplate disclaimers
📰 State-aligned narratives

These triggers aren’t hypothetical. They can appear accidentally (in intel docs) or be planted maliciously (via prompt injection).

[GovBench @GovBench]

How does @OpenAI's new GPT-5 perform on US Military domains? We ran it against our JointStaffBench(v1).

More about JointStaffBench: https://t.co/1TsaDVyWE7

[GovBench @GovBench]

Today we’re thrilled to announce that GovBench is incorporating as a 501(c)(6) nonprofit association.

Our mission is clear: close the gap between frontier AI research and day‑to‑day government work so agencies can deploy AI that is safe, accurate, and transparent.

Billions of taxpayer dollars are earmarked for AI this year alone. GovBench is setting the standards to ensure those investments deliver real public value.

Get Involved:
- Donate or partner: [email protected]
- Follow our journey: @GovBench on LinkedIn & X

Together, we can make AI work for government—and for everyone it serves.

https://t.co/fq84qxbrYh