
Tech • AI • Robotics • Game
Tests of AI-driven hacking show that closed frontier models from OpenAI and Anthropic remain the most capable, but cheap downloadable Chinese open-weight models are already good enough to automate broad cyberattacks at low cost.
In a recent cyber range test by Vals, AI models were given a workstation and a fake company target, then asked to penetrate networks, take control of machines and extract protected data without human help. A Chinese open-weight model, identified as DeepSeek, reportedly reached most objectives on a first network for under $3 per attempt. A leading OpenAI model also completed the objectives, but at roughly $20 per run.
Cybersecurity practitioners said the headline result should not obscure a broader reality: the strongest frontier models remain the closed systems run by US labs such as OpenAI and Anthropic. In more complex attack chains, those models are better at recovering from mistakes, adjusting strategy and using fewer actions. That matters because advanced intrusion work depends less on brute force than on maintaining coherent reasoning across many steps.
The strategic concern is not that open-weight models are universally better, but that they are accessible to anyone. Researchers described some Chinese models as effective “sprinters”: they can scan widely, try many combinations and uncover simple but real weaknesses at scale. Even if they consume more tokens and become easier to detect, they can still lower the cost of cybercrime enough to make small firms, local administrations and poorly defended industrial operators newly attractive targets.
Security engineers argued that the Vals evaluation used only two environments and scenarios that were too simple to reflect real enterprise networks. In more realistic tests with larger environments, success rates drop sharply, and the gap between models can widen. The criticism was not that the findings are useless, but that they may overstate how cheaply a model can compromise a meaningful target in practice.
Another test cited by the participants suggested that model rankings shift depending on spending. At very low budgets such as $0.10 or $1, smaller or cheaper models can offer the best return, with DeepSeek performing strongly early on. At $10 and especially $100, larger closed models regain the lead, showing stronger endurance on longer tasks. The pattern suggests that open models can quickly find low-hanging fruit, while frontier systems perform better over sustained operations.
The discussion stressed that many companies are far below the level assumed in frontier-model debates. Some small and midsize firms still lack basic internal security expertise, rely on minimal cyber insurance, and may not know how to test their own exposure. If the cost of launching attacks falls toward tens of dollars, the economics of extortion change: targets once considered too small to bother with may suddenly become worth attacking.
One major risk is the rise of inexperienced attackers using models as tutors and operators. A user can ask an agent to scan a site, identify a vulnerability and suggest next steps without understanding the underlying techniques. That does not eliminate the advantage of skilled operators, but it can expand the pool of people able to attempt intrusion, much as automation has lowered barriers in other online scams.
The broader context includes state-backed espionage and offensive cyber operations by powers including China, Iran, North Korea, the United States and European states. Isolated regimes may find cyber operations especially attractive because they require far less capital than conventional military programs. Western countries are widely seen as strong in intelligence and offensive cyber capabilities as well, even if they present themselves publicly as acting mainly in defense.
The debate extended to BYD and other connected Chinese vehicles, which combine cameras, microphones, geolocation and sometimes lidar. US forces reportedly barred such vehicles from Kadena Air Base in Japan, while Poland, the UK and Israel have also taken restrictive measures around sensitive sites. The concern is that mass-market connected products can become intelligence collection platforms, illustrating how economic globalization and digital infrastructure are colliding with national security.
The central issue is no longer whether AI can support hacking, but who can access the strongest systems and how fast cheap open models can expand the attacker base. Restricting frontier models may slow some misuse, yet it also risks leaving weaker defenders exposed while widely available tools continue to improve.
Ask a question