---
title: "AI Models Matching Top-Secret Counterparts in Cyber Capabilities"
url: https://newscentral.site/ai-models-matching-top-secret-counterparts-in-cyber-capabilities/
language: en
publisher: "News Central Site"
section: "Technology"
published: 2026-11-06T18:43:43.000Z
updated: 2026-11-07T06:54:19.148Z
id: 8fd88897-0e24-482c-b218-e1a05aec5b1b
source: "Wired https://www.wired.com/story/free-chinese-ai-models-are-getting-scary-good-at-hacking/"
attribution: "Link to https://newscentral.site/ai-models-matching-top-secret-counterparts-in-cyber-capabilities/ and name News Central Site when you quote or summarize this story."
---

# AI Models Matching Top-Secret Counterparts in Cyber Capabilities

Researchers at Vals have found that even open-source AI models can now match the capabilities of top-secret ones.

Researchers at Vals have been testing AI models to see how well they can carry out complex network instructions on mock corporate networks. The results show that even open-source models, available for anyone to download, are now matching the capabilities of top-secret models only accessible through trusted programs.

According to Nikil Ravi, a researcher at Vals, this shift is significant because it means that closed models no longer have an exclusive edge in cyber skills. This challenges recent efforts by top labs to limit access to their models and slow the spread of advanced hacking abilities.

The US frontier labs, particularly Anthropic, had previously released new models under limited access programs, starting with Claude Mythos Preview in April. These releases were based on the idea that gated access would only give hackers a temporary advantage before more powerful tools became widely available.

The transition from vulnerable AI models to more powerful ones has become a reality in the world of hacking. Vals, a cybersecurity research firm, has been testing advanced AI models in simulated scenarios to assess their capabilities.

In two separate tests, Vals presented AI models with specific tasks to achieve within mock corporate networks. The first scenario involved gaining access to a public-facing website that led to a company network. The second scenario required compromising one company network to gain access to another.

The results showed that OpenAI's Daybreak model, available only to trusted partners, was the most capable AI model in these tests. It completed 81 percent of its objectives on average, with some tasks showing equal performance from the Chinese open-weight model DeepSeek.

The use of open-weight AI models in cybersecurity is gaining traction, particularly among those on the defensive side. According to Rayan Krishnan, CEO of Vals, this trend provides a strong incentive for companies and developers to opt for open models over proprietary ones.

Much of the year's cybersecurity landscape has unfolded as predicted, with AI automating core hacking tasks to identify vulnerabilities in code and probe networks for weaknesses at an unprecedented speed and scale. The intensity of fixing uncovered software flaws has also increased significantly.

However, experts are still debating whether AI has fundamentally altered the conceptual foundation and tenets of cybersecurity defense. It's unclear whether AI tools pose a new kind of threat or amplify existing security challenges.

The fact that open-weight models are now matching their proprietary counterparts in terms of performance is significant for understanding how cybercriminals can harness AI, and how companies should respond to this development.

Even with strict controls in place, cybersecurity experts warn that defenders must assume top attackers have already had access to high-performance AI models developed behind closed doors.

Alex Zenla, cofounder and chief technology officer of cloud security firm Edera, believes open-weight models can already produce sophisticated attacks and vulnerabilities. However, he notes that these models may not be as advanced as those being secretly trained in closed labs, which could potentially surpass their capabilities once released.

The cat-and-mouse game between AI model developers and cybersecurity experts continues to escalate, with each side trying to outpace the other in a bid for security superiority.

---
Source: [Wired](https://www.wired.com/story/free-chinese-ai-models-are-getting-scary-good-at-hacking/)  
Published by News Central Site: https://newscentral.site/ai-models-matching-top-secret-counterparts-in-cyber-capabilities/