Key points
- DeepSeek completed 73% of simulated hacking tasks, nearing OpenAI's Daybreak score of 81%.
- Open-weight models are now nearly as capable as closed frontier models for network intrusion.
- Anthropic premised limited model releases on the idea that gated access would buy time.
Vals published research showing open-weight artificial intelligence models now match closed frontier systems in executing complex network intrusions. The study tested various AI systems on mock corporate networks to evaluate their ability to perform multi-step hacking instructions.Free AI tools can now hack networks almost as well as expensive, restricted ones.Vals data indicates open-weight models are converging with closed frontier systems in multi-step network exploitation capabilities. This finding signals a shift in the cybersecurity landscape where free, downloadable models compete with restricted, high-end cyber tools.
The claim
Researchers evaluated models using two scenarios involving simulated corporate networks. One test required accessing a network through a public website, while the other involved compromising one company network to reach a second. OpenAI’s Daybreak, a closed model for trusted partners, achieved the highest performance by completing 81 percent of assigned tasks.reported The open-weight Chinese model DeepSeek followed closely, completing 73 percent of objectives, with researchers noting equal scores on specific tasks.reported
The response
This development suggests that recent efforts by frontier labs to limit access to their models may be ineffective. The findings challenge the assumption that gated access provides a significant security advantage or slows the proliferation of advanced hacking capabilities.
Why this matters now
The results align with timelines previously suggested by industry leaders. Anthropic’s Project Glasswing, which provides trusted access to new model capabilities, was designed to prepare defenders. Logan Graham, Anthropic’s frontier red team lead, told WIRED in April that the company expected advanced hacking capabilities to become broadly available within 6 to 24 months.reported Seven months later, Vals’ data suggests this transition has arrived.
What buyers should weigh
The source material indicates that closed models no longer have a definitive edge in cyber skills. Wired reported that Anthropic premised the limited release of new models on the idea that gated access would only buy time before advanced hacking capabilities were broadly available.reported
Why it matters
- Open-weight AI models are now nearly as capable of hacking computer networks as the best closed models, signaling a turning point in cybersecurity.
- The finding that closed models no longer have a definitive edge suggests efforts by frontier labs to limit access may be moot.
- The rapid arrival of broadly available advanced hacking capabilities validates warnings that security landscapes would change significantly within months.
- Chinese open-weight model DeepSeek completed 73 percent of hacking tasks, performing not far behind OpenAI’s advanced cyber model Daybreak.
Topics: AI security, open-weight models, network intrusion, Vals research, Anthropic
Updated
The strongest doubt
Counterpoint from the sources
OpenAI’s Daybreak still achieved the highest completion rate, suggesting closed models retain a slight performance advantage in some scenarios.
For readers
Researchers
- Open-weight models now rival closed systems in multi-step network intrusion tasks.
- Gated access strategies may be less effective than previously assumed.
Policy
- Restricting model access may not significantly slow the spread of hacking capabilities.
- Defense preparation for broad capability availability is now urgent.
Investors
- The value proposition of proprietary cyber AI models may be eroding.
- Open-weight competitors are closing the performance gap rapidly.
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Which AI model performed best in the Vals hacking test?
OpenAI’s Daybreak completed 81 percent of tasks, while DeepSeek completed 73 percent.
Are open-weight models now as dangerous as closed models?
Vals research shows open-weight models are now almost as capable as the best closed ones in network intrusion.
Did Anthropic expect this outcome?
Yes, Anthropic stated that gated access would only buy time before these capabilities became broadly available.
Connected stories
Reading receipt
- DeepSeek completed 73% of simulated hacking tasks, nearing OpenAI's Daybreak score of 81%.
- Open-weight models are now nearly as capable as closed frontier models for network intrusion.
- Anthropic premised limited model releases on the idea that gated access would buy time.
Still open: How will this shift impact the pricing and strategy of proprietary cyber AI vendors?
Still open: Will open-weight models continue to close the remaining performance gap with Daybreak?
