Source time Edition date
IndustryWired AI
Free Chinese AI Models Are Getting Scary Good at Hacking
免费中国AI模型在黑客攻击上强得吓人
DeepSeek and GLM matching OpenAI's top cyber model in hacking tests raises a dilemma: defenders gain cheap tools, but attackers do too, and no one agrees who should gate access.
Judgment
CISOs should decide whether to standardize on open-weight models for vulnerability hunting at $2 to $3 per attack versus $17 to $46 for GPT-6-Sol, accepting that attackers can download the same capability; the tradeoff is cheaper defense at the cost of broader proliferation.
Basis
Vals tested AI models on two simulated corporate networks. Nikil Ravi said the best closed frontier models "are now starting to be matched by open models that anybody can download." OpenAI's Daybreak completed 81 percent of tasks on average; DeepSeek completed 73 percent, with some tasks tied. Vals says GPT-6-Sol attack tokens cost $17 to $46 versus $2 to $3 for DeepSeek, because open models can run on own hardware. CEO Rayan Krishnan says that cost gap incentivizes defenders to use open models. The mechanism: if open weights match closed cyber capability, gated release only delays access; the same model weights can be downloaded and run locally, so defenders and attackers face similar capability at lower cost. Anthropic's Logan Graham had argued gated access buys time for defenders; Alex Zenla cautions closed labs may have unreleased models exceeding open ones and calls it a cat-and-mouse game.
- Daybreak task completion81 percent
- DeepSeek task completion73 percent
- GPT-6-Sol attack cost$17 to $46
- DeepSeek attack cost$2 to $3
Unclear The source does not say whether the tested open-weight models had safety training disabled (abliterated), nor how often real attackers used DeepSeek or GLM rather than older models.
Next Watch whether Vals or another evaluator publishes a follow-up test that includes abliterated open-weight models and real-world incident data, so the parity claim can be checked against attacker behavior.