UK AISI and U.S. CAISI Publish Preliminary Assessment of Kimi K3’s Cyber Capabilities

Jul 23, 2026

A preliminary joint assessment by the UK Artificial Intelligence Security Institute and the U.S. Center for AI Standards and Innovation found that Kimi K3 performed significantly below the most recent frontier cyber-capable models on preliminary cyber evaluations while outperforming GLM-5.2. Kimi K3 achieved arbitrary code execution on 0 of 41 ExploitBench samples. Separately, White House official Michael Kratsios accused Moonshot AI of distilling Anthropic’s Fable model to develop Kimi K3. U.S. Treasury Secretary Scott Bessent said the United States could sanction overseas AI models if they were found to be stealing from American companies, and the Trump administration was reportedly considering a ban on U.S. companies hosting Chinese AI models unless their security could be guaranteed.

Kimi K3 AI model release

▪Kimi K3 has 2.8 trillion parameters and its weights occupy 1.4 terabytes of storage.
▪Moonshot AI temporarily paused new subscriptions for Kimi K3 within two days of its launch due to high demand.
▪Moonshot AI released the Kimi K3 artificial intelligence model on July 16, 2026, with an open-weight release scheduled by July 27, 2026.

US-China AI competition

▪On the ExploitBench benchmark, Kimi K3 scored 32%, outperforming China's GLM-5.2 at 24% but lagging behind top U.S. models which averaged 76.2%.
▪A joint evaluation by the UK Artificial Intelligence Security Institute and the U.S. Center for AI Standards and Innovation found that Kimi K3's cyber capabilities trail leading American frontier models.
▪Kimi K3 failed to achieve arbitrary code execution on all 41 ExploitBench tasks, whereas leading U.S. models achieved it on an average of 20 tasks.
▪On "The Last Ones" cyber range, Kimi K3 reached step 17 of 32 on average, while leading U.S. models reached 28.5 steps on average.

Trump administration AI policy framework

▪President Donald Trump stated on July 23, 2026, that the United States must dominate artificial intelligence and cryptocurrency to remain the number one superpower.
▪The Trump administration executive order gave federal agencies until August 1, 2026, to establish a classified benchmarking process for covered frontier AI models.
▪President Donald Trump signed an executive order in early June 2026 establishing a voluntary 30-day pre-release testing framework for frontier AI models.

Industry responses to Chinese AI

▪OpenAI Chief Global Affairs Officer Chris LeHane stated that the release of Kimi K3 reinforces the need for the U.S. to establish a coherent AI testing framework.
▪Venture capitalist David Sacks argued that U.S. regulations, data center bans, and proposed federal pre-approval agencies are risking America's competitive edge in AI.
▪Tech investor Chamath Palihapitiya criticized using a “China boogeyman” to protect frontier labs’ business model and said the market should sort it out.

Policy responses to Chinese AI

▪Anthropic Head of Public Policy Sarah Heck characterized illicit adversarial distillation as intellectual property theft and industrial espionage that poses national security risks.
▪The Trump administration is reportedly considering a ban on U.S. companies hosting Chinese AI models unless their security can be guaranteed.
▪White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot AI of cloning Anthropic's Fable model capabilities through large-scale distillation.
▪U.S. Treasury Secretary Scott Bessent stated that the United States has the ability to sanction overseas AI models if they are found to be stealing from American companies.