Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4
Automated AI safety research and reduced-safeguard open-weight models raise concrete questions about the pace and reliability of AI governance mechanisms - relevant for agencies tracking AI risk.
Key points
- Anthropic researchers show AI agents can automate alignment research, outperforming humans on a weak-to-strong supervision benchmark.
- A safety evaluation of Chinese open-weight model Kimi K2.5 finds fewer CBRN refusals and greater misaligned behaviour than Western frontier models.
- Huawei's HiFloat4 training format outperforms the Western MXFP4 standard on Ascend chips, reflecting export-control-driven efficiency pressure.
Implications for Australian agencies
- Monitor AI safety and risk teams may want to monitor Anthropic's automated alignment research programme as an early signal of how quickly AI safety R&D itself could be accelerated or destabilised.
- Monitor Agencies assessing AI procurement risk or CBRN-adjacent use cases could note the Kimi K2.5 findings, particularly the low cost of safeguard removal in open-weight models.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 20 April 2026
"Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4"
Source: Import AI – Substack (Jack Clark)
Published: 20 April 2026
URL: https://importai.substack.com/p/import-ai-454-automating-alignment
Import AI issue 454 covers three substantive research developments. Anthropic demonstrates that Claude-based automated agents can conduct alignment research end-to-end, achieving a 0.97 performance gap recovered score versus a 0.23 human baseline on weak-to-strong supervision - though results did not generalise to production models. An independent safety evaluation of Kimi K2.5, a leading Chinese open-weight model, finds it has CBRN refusal rates significantly lower than GPT and Claude equivalents, with safeguards strippable for under USD $500 in compute. Huawei's HiFloat4 format outperforms the Western MXFP4 standard on its own Ascend chips, illustrating how export controls are accelerating Chinese hardware-software co-optimisation. The issue also includes a brief on Ukraine's first fully robotic battlefield victory and a fictional short story.
Implications for Australian agencies:
- [Monitor] AI safety and risk teams may want to monitor Anthropic's automated alignment research programme as an early signal of how quickly AI safety R&D itself could be accelerated or destabilised.
- [Monitor] Agencies assessing AI procurement risk or CBRN-adjacent use cases could note the Kimi K2.5 findings, particularly the low cost of safeguard removal in open-weight models.
Retrieved from SIMS, 18 July 2026.