Open-Weight AI Models Reach Parity With Closed Systems in Hacking Tasks
New testing reveals that open-source AI models are nearly as capable of network intrusion as top closed models, shifting the cybersecurity landscape.
Recent research from AI evaluation firm Vals indicates that open-weight artificial intelligence models are approaching the cyber capabilities of the most advanced proprietary systems. According to Nikil Ravi, a researcher at Vals, the gap between gated frontier models and downloadable open models is narrowing, suggesting that strict access limitations by labs may no longer effectively slow the spread of these technical skills.
The findings challenge assumptions held by US frontier labs. Earlier this year, companies like Anthropic argued that restricting model access would buy time for defenders. For instance, Anthropic's Logan Graham noted in April that restricted access programs like Project Glasswing aimed to give defenders a head start before advanced capabilities became widespread. Seven months later, however, open models are rapidly catching up.
Performance and Cost Comparison in Simulated Tests
Vals evaluated various artificial intelligence models across two simulated corporate network environments. The first involved a public-facing corporate network, while the second required models to breach one company network to gain access to a second.
Among proprietary systems, OpenAI's advanced cyber model, Daybreak—accessible only to trusted partners—completed 81 percent of assigned tasks across tests. Meanwhile, the Chinese open-weight model DeepSeek achieved 73 percent, matching the top closed model on several specific challenges.
Beyond performance, the research highlighted a stark contrast in operational costs. Running an attack via proprietary models like GPT-6-Sol costs between $17 and $46 in tokens, whereas utilizing DeepSeek for similar tasks ranges from $2 to $3. Rayan Krishnan, CEO of Vals, points out that this cost efficiency offers a compelling incentive for defensive teams to leverage open models.
What it means for developers
For software engineers and security professionals, this convergence of capabilities changes how both defensive and offensive automation will be deployed. While open models provide affordable alternatives for security analysis, they also lower the barrier for malicious actors. Developers can experiment with and integrate top AI models cheaply through one API at https://apixoai.online. Despite the advanced skill sets demonstrated by models like DeepSeek, real-world observations indicate that malicious actors often rely on older, less capable models to conserve resources.
Alex Zenla, cofounder and CTO of cloud security firm Edera, notes that open-weight models can generate advanced attacks, particularly when safety training is modified. However, Zenla points out that closed labs likely retain unreleased models exceeding current open capabilities, maintaining a continuous cycle between offense and defense.
Source: Free Chinese AI Models Are Getting Scary Good at Hacking — Wired AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
