UK AISI cyber evals seem to under-elicit capabilities from quirky models [1]. Kimi K3 is a token-hungry model, and I suspect it hit the eval's 100M token limit well before saturating scores [2].

This gap was true for GLM 5.2 as well; they ranked it at Opus 4.5 level [3]. Both anecdotally and with a held-out eval, I've found GLM 5.2 to be better at security research than Opus 4.6 [4]. But it's a quirky model that degrades quickly at long context lengths.

Personally, I'd rank Kimi K3 above Opus 4.8 and lower than GPT 5.6 Sol in its ability to find vulnerabilities and exploit them. But it's not far from the frontier.

[1]: From the UK AISI: "Our setup likely slightly underestimates open weight models’ maximum capability: we didn’t pursue specific elicitation or optimisations which could have improved performance" (https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are...).

[2]: Their eval also counts cache hits towards the token budget; the 100M token budget is comparable to a ~5M token budget in other evals.

[3]: See GLM 5.2 eval scores in https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are...

[4]: https://dualuse.dev/posts/chinese-models-are-sometimes-bette...

Just looking at the Elo differences, k3 is at ~2000 Elo, compared to SotA closed models at 3000 Elo. That is a huge difference. Also, even on the public benchmarks, k3 only scores in the "low hanging fruit tasks", with 0 successful code execution or arbitrary r/rw scenarios.

> Kimi K3 achieved ACE on 0/41 samples, whereas the most cyber-capable models achieved ACE on 20/41 samples on average

But there's hope:

> Kimi K3 reached step 17 of this 32-step attack path on average, while the most cyber-capable U.S. models reached 28.5 steps on average.

> In one of the 10 attempts, Kimi K3 successfully completes “The Last Ones” cyber range within the 100M token limit. This indicates that Kimi K3 is capable of autonomously attacking small, weakly defended and vulnerable enterprise systems, when directed to do so and given initial network access. [...] the most capable models solving it more reliably at 6/10 and 7/10 attempts

The bigger problem is not raw capability, IMO. That can be further RLd into surfacing more reliably. The bigger problem, as seen in the HuggingFace scenario is that SotA models might hit classifiers / guardrails randomly, and leave you with plain refusals. In that case, it is probably better to have something that can help, locally, rather than rolling the dice with API based systems that are more capable but can just refuse arbitrarily.

Anyway, one of the lessons here is to take with a grain of salt every "x model has caught up with ySotA model". They likely haven't for the breadth of tasks that SotA can handle today. Also, number goes up on benchmarks has been a thing for years, and every time a new (or closed one) appears, the gaps are again obvious (and large).

Also, the key benefit of open models is that whatever capabilities they reach, those will never go away. You will always be able to access them, at that level, going forward. Investing in running those models gives you stability, and you don't suddenly lose a capability because API provider x decided to sunset a model family.

For the purposes of model selection, it's irrelevant to an attacker if a model achieves an offensive objective 70% of the time, when that model refuses to participate 100% of the time. However, a model that always participates but "only" succeeds 10% of the time is golden -- just run it more often, or give it more tokens. Attackers are patient, and many of them are well-resourced.

Am I reading this wrong?

Even if we know what the set is, it's not clear what the numbers actually mean. Is it min, max, average, weighted, median of the models? Prerelease or public, with or without safeguards?

A bit frustrating to have this be hand waved in a report, the graphs might as well just have two mystery bars, U.S. and China.

May be the case, but also their other graph shows that these open Chinese models are only about 6 months behind the frontier. Plus normal people are allowed to use them. So if Mythos has genuinely scary hacking capabilities (which seems to be the case), then we should expect that in open weights models early next year.

I think at this point if you aren't constantly pointing a frontier model (or Kimi K3 if you aren't well connected) at your infrastructure / code and asking it to hack you then you're being negligent.

these two claims can't be true at one and the same time:

(a) they're distilling our secret sauce!

(b) they'll never catch us!