Open-Weight AI Models Now Match Closed Frontier Systems at Hacking Tasks
Research from Vals shows open-source Chinese models like DeepSeek are closing the cybersecurity capability gap with proprietary systems, undermining access-control strategies.
Last verified:
Open Models Narrow the Hacking-Capability Gap
According to Wired AI, research firm Vals has documented that open-weight AI models now perform nearly as well as proprietary frontier systems on complex network penetration tasks. DeepSeek, a freely downloadable Chinese model, achieved a 73% success rate on a battery of multi-step hacking objectives, trailing only OpenAI’s restricted-access Daybreak model at 81%. The finding marks a critical inflection point: the performance gap that justified gated access to frontier capabilities has narrowed to single-digit percentage points in just seven months—far shorter than the 24-month window Anthropic publicly estimated in April.
Vals structured its testing around two scenarios: first, attacking a simulated corporate network through a public-facing web interface, and second, compromising one network to pivot into a second. The researchers ran models through multiple test iterations, measuring both success rates and token consumption costs. The results show that while OpenAI’s proprietary system retains a marginal lead, the difference is not decisive. On some individual tasks, DeepSeek and Daybreak scored identically, according to Wired AI.
Cost Asymmetry Favors Open-Weight Deployment
The economic implications may prove more consequential than raw capability parity. Wired AI reports that executing an attack using OpenAI’s GPT-6-Sol costs $17–$46 in API token fees, whereas the same task costs $2–$3 when run on DeepSeek using local compute. This 8–23× cost reduction inverts the traditional defense economics: defenders now face financial incentives to deploy open models rather than subscribe to expensive proprietary APIs. Vals CEO Rayan Krishnan, according to Wired AI, explicitly flagged this reversal as a factor tilting the advantage toward attackers with access to commodity hardware.
Why This Matters
Anthropic’s April strategy, confirmed by the company’s frontier red team lead Logan Graham in an interview with Wired, centered on buying time for defenders through restricted model access. The company framed limited release as a mechanism to compress the transition from proprietary advantage to broad availability into a window long enough for security teams to prepare mitigations. Wired AI’s coverage of Vals’ findings suggests that window may have closed ahead of schedule. The convergence of open-model capability with closed-model performance—coupled with the cost advantage of running models locally—removes a key lever that frontier labs believed would slow the proliferation of AI-automated hacking. Organizations relying on the assumption that advanced attack automation would remain scarce and expensive for months to come may need to accelerate both defensive tooling and incident-response preparation.
Frequently Asked Questions
What specific hacking tasks were the models tested on?
According to Wired AI, Vals tested models on two scenarios: compromising a mock corporate network via a public-facing website, and laterally moving from one compromised network to access a second. Tasks were multi-step, network-level operations.
How much does it cost to run these attacks with each model?
Wired AI reports OpenAI's GPT-6-Sol costs $17–$46 per attack in token costs, while DeepSeek costs $2–$3, making open models 8–23× cheaper to operate.
Why does this matter for cybersecurity defense?
Vals CEO Rayan Krishnan notes the cost advantage incentivizes defenders to use open models rather than proprietary APIs, but the parity in capability means enterprises cannot rely on scarcity to slow attack automation.
Did Anthropic anticipate this timeline?
Yes. Anthropic's frontier red team lead Logan Graham told Wired in April that the company expected advanced hacking capabilities to be 'broadly available in 6, 12, 24 months.' The Vals findings suggest the shortest timeline is the operative one.