Open models close the cyber gap, Kimi K3 rises, and Demis maps AGI regulation
Import AI (Jack Clark) · rss · 2026-07-20
This issue of Import AI covers three major threads.
1) The open-vs-closed cyber gap is narrowing
The UK AI Security Institute analyzed cybersecurity capabilities across leading proprietary and open-weight models. Its conclusion: the gap has shrunk this year. On 70 narrow cyber evals, recent open models such as GLM-5.2 and DeepSeek V4-Pro now sit much closer to frontier closed models than before. The gap is larger on long-horizon cyberrange tasks, suggesting open models are still weaker at chaining capabilities into full attack workflows.
The policy implication is urgent: if powerful cyber capabilities diffuse through open-weight releases, defenders have a short window to prepare before those capabilities become widely available without proprietary safeguards.
2) Kimi K3 signals stronger Chinese frontier models
The newsletter also highlights Kimi K3, a 2.8T-parameter model that reportedly performs at frontier level across many benchmark-style tasks, sometimes matching or trailing the best proprietary systems. The write-up notes brittleness that looks like benchmark overfitting, but also emphasizes how powerful widely available models could fuel entrepreneurship and broader access to compute-like “sovereign intelligence.”
It also mentions example use cases around recursive self-improvement: Kimi K3 wrote a compact Triton-like GPU compiler and, in another demo, designed and verified a chip in a 48-hour autonomous run using open-source EDA tools.
3) Demis Hassabis proposes a frontier AI regulatory regime
Demis Hassabis argues for a US-led standards body, modeled loosely on FINRA, to test frontier AI systems before release. The body would define frontier-model criteria, run assessments with federal agencies and national labs, and encourage labs to share models for review up to 30 days before launch. The broader idea is to move from voluntary review toward formal regulation once the testing process proves effective.
Overall, the issue frames frontier AI as a three-way story: cyber capability diffusion, Chinese model progress, and emerging regulatory infrastructure.
More from Models
- Dev benchmarking skills finds Sol at 17 tok/s vs Opus 5.5 at 70 tok/s — lucasmeijer · 2026-10-26
- Fern's "context is intelligence" thesis: prediction keeps improving out to 10.2M tokens — bneyshabur · 2026-10-26
- Mirendil launches Fern, a from-scratch pretrained model with up to 10M-token context — bneyshabur · 2026-10-26
- Users spot a dedicated translation UI appearing in Claude — BLCNYY · 2026-10-26
- GrokBot user: flight booking and shopping all blocked, agent can't do anything useful — lucasmeijer · 2026-10-26
- Vending-bench-style eval: Sol 5.6 appears the most aggressive contract-violator — xuanalogue · 2026-10-26