Free Open Models Are Closing the Gap on Costly Cyberattack Tools

The UK's AI Security Institute found open-weight models now trail top proprietary systems by four to seven months on cyber tasks — at a fraction of the cost.

The British AI Security Institute assessed how far open-weight AI models lag proprietary systems in cyber capabilities and found the gap has narrowed to four to seven months, down from six to ten months for most of 2025. GLM-5.2, released in June 2026, matched February's Opus 4.6 on narrow cyber tasks, while DeepSeek V4-Pro performed at the level of Opus 4.5 from November 2025.

The cost difference is the sharp part. On a 100-million-token Cyber Range test, Opus ran about $85, GLM-5.2 about $46, and DeepSeek V4-Pro about $1.19. Per-task costs told the same story: roughly $15 for Opus 4.6 versus 28 cents for DeepSeek V4-Pro. AISI also found the open models' safety measures were largely ineffective and could be bypassed simply by retrying. Benchmarks spanned 70 narrow tasks and multi-step network attack simulations.

We file this as friction. Cheaper, more capable, openly available models mean the ability to research vulnerabilities and run exploits gets broadly accessible. That lowers the barrier to offensive hacking, gating the security and trust that abundance depends on.

The caveats matter. AISI treats the Cyber Range results as weaker evidence, the tests omit real-world defenders, and they can't always separate capability gaps from failures at long, complex planning. It's also unclear whether future open models will match recent closed-model gains. AISI plans to test Kimi-K3 next, with weights due late July.

Source: The Decoder