AI
Open-Source AI Models Now Match Proprietary Systems at Network Hacking
Freely available AI models have caught up to expensive closed systems in their ability to compromise corporate networks, complicating efforts by labs to control dangerous capabilities.
WIRED — AI
The gap between restricted, proprietary AI and freely downloadable models is shrinking fast when it comes to cybersecurity threats. New testing reveals that openly available AI systems can now perform sophisticated network intrusions nearly as well as the most advanced closed models—a threshold that cybersecurity experts have long anticipated but hoped to delay.
What the Research Shows#
Vals, a firm that benchmarks AI capabilities, ran a controlled experiment pitting different language models against simulated corporate networks with escalating difficulty. In their first test scenario, models faced a mock company infrastructure exposed through a public website. The second scenario required compromising one network to access a second, more complex target.
OpenAI's Daybreak—available only to trusted security researchers—completed 81 percent of assigned hacking tasks. DeepSeek, an open-weight Chinese model anyone can download, achieved 73 percent. On individual tasks, the two models sometimes performed identically. "The best frontier models, locked up inside the labs and available only through trusted access programs, are now starting to be matched by open models that anybody can download," according to researcher Nikil Rav at Vals.
The Economics of Hacking#
A key advantage of open models lies in cost. Running DeepSeek for an attack costs $2 to $3 in tokens versus $17 to $46 for proprietary alternatives. Because open models run on users' own hardware rather than through cloud providers, they're far cheaper to operate at scale—creating a financial incentive even for defenders to switch away from paid services.
Why Labs Restricted Access#
Leading AI companies, particularly Anthropic, justified their controlled-release strategy by arguing that limiting access would buy time before dangerous capabilities proliferated. Anthropic's red-team lead told WIRED in April that restricted access would help defenders "get a head start" before capabilities became widely available in 6 to 24 months. That timeline appears to have been accurate.
The Broader Security Puzzle#
The findings complicate an ongoing debate about whether AI truly reshapes cybersecurity fundamentals or simply amplifies longstanding vulnerabilities. Throughout 2024, AI tools have accelerated both attack and defense—automating vulnerability discovery and network reconnaissance at previously impossible speeds. Yet real-world cybercriminals have been slow to adopt cutting-edge models, typically relying on older, less capable systems to conserve resources.

