Free Chinese AI Models Are Getting Scary Good at Hacking

The cybersecurity landscape is undergoing a profound transformation, with the emergence of freely available AI models rapidly closing the gap with their more restricted counterparts in their ability to penetrate computer networks. A recent study from Vals, a company specializing in AI capability measurement, reveals that open-weight AI models are now approaching the hacking proficiency of the most advanced, closely guarded systems – a development that casts doubt on strategies aimed at controlling access to these powerful tools. This shift carries significant implications for businesses, government agencies, and individual residents in New York City and beyond, as the potential for increasingly sophisticated cyberattacks intensifies.

Vals’ research directly challenges the prevailing assumptions of companies like Anthropic, which recently released its Claude Mythos Preview model with limited access, hoping to buy time before advanced AI hacking capabilities became widely available. Logan Graham, Anthropic's frontier red team lead, previously stated that providing trusted parties with early access was intended to “give a head start” to defenders. “We need to prepare now for a world where these capabilities are broadly available in 6, 12, 24 months. Many things would be different about security,” he told *WIRED* at the time. This perspective reflected a broader belief within the industry that controlled release could mitigate the risks associated with powerful AI. Seven months later, however, Vals’ findings suggest that timeline may have been overly optimistic, and that the democratization of AI hacking tools is happening more quickly than anticipated.

Vals' Methodology and Findings

The Vals research involved testing AI models in two simulated scenarios designed to mimic real-world corporate network environments. The first scenario involved a mock corporate network accessible through a public-facing website, representing a common entry point for attackers. The second, more complex scenario required the model to compromise one company network in order to gain access to a second, simulating a multi-stage attack – a tactic frequently employed by sophisticated threat actors.

Model Performance Comparison

The company found OpenAI’s most advanced cyber model, Daybreak—available only to trusted partners—to be the most capable, completing 81 percent of the tasks it was given in an average of two tests. The Chinese open-weight model DeepSeek was not far behind, completing 73 percent of the objectives on average. On some tasks, Vals researchers noted that the two models scored equally. This parity is particularly striking when considering the cost differential: using DeepSeek costs significantly less; Vals estimates between $2 and $3 per task compared to $17 to $46 for GPT-6-Sol, a significant incentive for defensive use. The economic advantage of open-weight models further underscores their potential for widespread adoption.

Implications for Cybersecurity Strategies

This shift is significant because it undermines the effectiveness of strategies that rely on restricting access to advanced AI models as a primary means of cybersecurity defense. The findings also highlight a crucial economic factor: open models are generally cheaper because they can be run on one's own hardware, reducing reliance on expensive cloud computing resources. Rayan Krishnan, CEO of Vals, emphasizes this as a strong incentive for those playing defense to use open models rather than proprietary ones – a move that could potentially level the playing field between attackers and defenders. For more information on Vals’ approach to AI capability measurement, visit Visit Vals website for more information.

Community Response and the Cat-and-Mouse Game

The cybersecurity community is taking note of these findings. Alex Zenla, cofounder and chief technology officer of the cloud security firm Edera, warns that this may be the beginning of a dangerous security landscape. “I am in general agreement that open-weight models are capable of producing advanced attacks and vulnerabilities today,” says Zenla. “My caveat to this is that it is likely that the closed labs have trained additional, unreleased models exceeding these capabilities. So I would put this very much like a cat-and-mouse game.” He believes labs are constantly striving to outpace open-weight models by securing vulnerabilities before they are found, creating an ongoing cycle of innovation and counter-innovation.

The Rise of AI Automation in Cybersecurity

The rise of open-weight AI hacking tools is occurring against a backdrop of increasing automation within the cybersecurity field. Throughout the year, AI has been increasingly employed to automate core hacking tasks, assisting both attackers and defenders in identifying vulnerabilities and probing networks. While the speed and scale of vulnerability discovery have increased, experts remain divided on whether AI has fundamentally altered the underlying principles of cybersecurity defense. This debate centers on whether AI represents a new kind of threat or simply amplifies existing security challenges. The Vals findings contribute to this discussion by demonstrating that open-source AI is rapidly becoming a viable tool for sophisticated attacks.

Pragmatism in Cybercrime and Future Concerns

While the findings suggest a concerning trend, real-world observation has shown that attackers often conserve resources and use older, less capable models—both proprietary and open—for most of their activity. This pragmatic approach reflects the high cost of cybercrime investigations and the need to maximize return on investment. Despite this, defenders must assume that the best-funded and most highly motivated attackers have already had at least some access to closed models all this time, highlighting the importance of proactive defense strategies. The research also acknowledges that “abliterated models” – those with disabled safety training – pose a significant risk, as they can be used to circumvent ethical guidelines and safety protocols.

Looking Ahead: Vigilance and Proactive Measures

The shift in AI capabilities is a stark reminder of the rapid pace of technological change and its implications for security. As these models become more capable, it is essential for families and businesses in New York City to stay vigilant and proactive. The digital infrastructure underpinning our daily lives—from financial transactions to critical infrastructure—is increasingly vulnerable to attack. By staying informed and prepared, residents can help ensure that their families and communities remain protected in an increasingly complex world. To understand how AI is automating core hacking tasks, read this article on *WIRED* AI Automates Hacking Tasks.

For more information on the latest developments in cybersecurity, check out these resources: WIRED's Cybersecurity section and The Electronic Frontier Foundation.

Disclaimer: This article is for informational purposes only and does not constitute cybersecurity advice. Consult with a qualified professional for specific security recommendations.

External Links: Read the full article on Wired, Visit Vals website for more information, Learn more about Anthropic's AI research, AI Automates Hacking Tasks, WIRED's Cybersecurity section, The Electronic Frontier Foundation

Source: *WIRED*

Reporter: Marcus Chen

Date: October 26, 2023