Most cyberattacks on record stem from closed models, not open-source ones
Public evidence shows most cyberattacks to date stem from closed models, not open-source ones; by the cyber-risk logic used to justify banning open models, you must also ban closed models' public APIs—otherwise, you only widen the offense-defense gap.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Three camps in the cybersecurity debate have never truly crossed paths
The author divides the open-model security debate into three camps: US frontier labs and national-security circles who view open-weight models as unacceptable social risk; a moderate camp including the author, Hugging Face, and Joshua Saxe who argue that banning open models makes the world less safe; and Chinese companies who publish open-source models with strong network capabilities based on their own risk assessments. The risk-maximalist voice dominates Western discourse, while substantive discussion of how Chinese labs evaluate risk barely exists—replaced mostly by claims like "Chinese labs don't care about safety," which border on ad hominem.
Anthropic's GLM-5.3 test sidesteps the policy questions that matter
Anthropic's latest report treats GLM-5.3 as a hostile cyberweapon test. The author sees the technical research as valid but faults the report for dodging the crucial cross-domain questions: what happens if we actually ban open models over cyber risk, and why did Chinese labs decide releasing these models was safe? The author describes this technical-only framing as "solipsistic"—self-talking-to-itself—when all policy moves are actually tradeoffs within a multi-party game, and rarely feature one universally superior choice.
Public evidence shows most cyberattacks to date came from closed models
Some argue (citing classified briefings) that one particular Chinese open model must be banned now, but the author points out that by public record, most documented cyberattacks have actually been caused by closed models (including broader FelonyBench data). This undercuts the popular claim that "open is dangerous, closed is safe"—the truer picture may be "open is unsafe, closed is also unsafe." Closed models are more capable and theoretically more guarded, but when those guardrails leak like sieves, greater capability could actually inflict greater net harm.
If you ban open models for cyber risk, you must ban closed APIs too
The author draws a logical conclusion: if the latest open models must be prohibited to slow cyber-risk diffusion, the same logic demands that frontier closed models' public APIs be banned as well. Closed models do have more guardrail layers than open models, but far from perfect ones; cyberattack capability may grow far faster than guardrail capability improves. Banning only open models while letting closed-model progress continue only widens the offense-defense gap. In the near term, open models may be the only tools deployable to air-gapped networks in national-defense settings.
China's AI-safety logic prioritizes political stability, not counterbalancing bad AI with good AI
The author believes China does care about AI safety, but its framework is native to Chinese culture and power structures, centered on political stability rather than the Western model of "good AI countering bad AI." Chinese companies must register major models with the government and submit evaluations—a system originally designed for information control, not cyber or biorisks. China's personal risk is far higher than America's: elite AI researchers cannot travel freely, and the sector faces tightening caps on foreign investment. By contrast, the worst American outcome is merely corporate dissolution.
Full safety audit for a frontier model like Kimi K3 costs tens of millions
Comprehensive safety evaluation of a frontier model at Kimi K3's scale may cost tens of millions of dollars in compute, money Chinese labs would rather invest in training. The real question isn't "do Chinese labs skip safety testing entirely" but "what's the minimum compute a lab should spend on safety before release?" That floor should exceed what Chinese labs currently spend, but probably needn't match Anthropic or OpenAI's level.
No amount of "using AI for good" plugs your own security holes
Western labs' core faith resembles "the best way to stop bad actors using AI is for us to use AI well," but the author neither trusts these institutions wholesale nor fully trusts their security teams individually. He cites accidents between Hugging Face and OpenAI, and the Hacktron breach of OpenAI's systems: if your own house is full of leaks, you're a poor candidate for the world's trusted AI-security partner.
If Claude Mythos had leaked as open weights, the world would probably be fine
Claude Mythos was billed as a novel cyberweapon that, if seized, would cause societal catastrophe. Instead, GLM-5.3 crossed the threshold first, and more than a month after its open-weight release there's scant evidence the world has materially changed. The author goes further: had Claude Mythos accidentally leaked in open weights back then, the world likely would have been okay—at worst, a visible bump in cyber incidents, closer to risk acceleration than phase shift. The author sees this as a falsifiable prediction by open-model doomers, one likely to be proven wrong.
In their own words · checked verbatim
The trope "Open Dangerous, Closed Safe" may be closer to "Open Unsafe, Closed Unsafe."
if you think the latest open-weight models need to be banned to slow the diffusion of cyber risks, you probably also need to make public-facing APIs for the frontier closed-models illegal.
It is very likely that cyber capabilities of closed models increase much faster than guardrail performance, and a world where open models are banned while closed models continue to progress would be increasing the offense-defense cyber gap.
China cares about AI safety. They have come to this discourse in a complete, endogenous manner relating to Chinese culture and power structures.
Running comprehensive safety evaluations on a frontier model like Kimi K3 could cost tens of millions of dollars in compute.
it’s hard to be the world’s trust cyber partner if your own house is teeming with leaks.
The leading proponents of this new wave of open-weight fear mongering have effectively been making a falsifiable prediction — that the current open-weight models are going to cause substantial, not seen before AI harms by crippling our cyber infrastructure.
If Claude Mythos was accidentally released as open-weight, it seems like the world would have been more or less fine.
Figures
| Full-scale safety evaluation cost for models like Kimi K3 | Tens of millions of dollars in compute | 1:34 |
| Time since GLM-5.3 open-weight release | More than one month | 1:34 |
Glossary
- open-weight models
- Models with publicly released weights that anyone can download and run locally.
- offense-defense gap
- The growing disparity when attack capability grows faster than defensive or guardrail capability improves.
- air-gapped networks
- Networks physically disconnected from external connections with no internet access; high-security closed systems.
- FelonyBench
- A benchmark aggregating documented AI-related cybercrime and misuse cases.
How to listen
Policy researchers and AI practitioners tracking US-China AI competition, open-source model regulation, and cybersecurity-risk debates.
Closing subscription reminder (from 9:04) can be skipped.