What happened
Cybersecurity firm Aikido reported that DeepSeek V4 Pro, an open- model, outperformed Google's Gemini 4 Argon and other closed models in identifying software vulnerabilities. While Google initially restricted Argon to vetted defenders due to cyberattack risks, it reached general availability two weeks later. Aikido's tests on 32 disclosed CVEs showed DeepSeek V4 Pro rediscovered 28 vulnerabilities (87.5% recall) at a total cost of approximately $295, whereas closed models like Anthropic's Opus 5 and Alibaba's Qwen 3.8 achieved 81.3% recall at higher individual costs.
On September 30, Google announced Gemini 4 Argon, its most capable model, but initially restricted access to a small group of vetted cybersecurity defenders through the Fairwind Program. Google DeepMind's chief AI architect, Koray Kavukcuoglu, stated that safely releasing frontier capabilities required a phased approach, citing concerns that the model's ability to find software vulnerabilities could be misused for cyberattacks.
Three weeks later, cybersecurity firm Aikido published results from testing ten AI models against 32 freshly disclosed CVEs. The test focused on genuine vulnerability reasoning rather than memorized answers. DeepSeek V4 Pro, an open- model available for public download, topped every public closed model on pooled vulnerability recall.
DeepSeek V4 Pro rediscovered 28 of the 32 vulnerabilities, achieving an 87.5% hit rate. This outperformed Anthropic's Opus 5 and Alibaba's Qwen 3.8, which both achieved 81.3% recall. The cost difference was significant: three runs of DeepSeek Pro cost Aikido roughly $295, while a single pass from Opus 5, Grok 4.6, or OpenAI's Sol cost between $450 and $590 and found fewer bugs.
Google's initial restriction on Gemini 4 Argon was not permanent. The model reached general availability on October 13, just two weeks after the restricted rollout began. It became available on the paid tier of Google's API, to Google AI Pro and Ultra subscribers in the Gemini app and Search, and in Google Antigravity for Ultra subscribers.
Other open- models also showed strong performance in Aikido's results. Qwen, Kimi, and Zhipu's GLM-5.3 demonstrated consistent vulnerability-finding capabilities. In a separate Cyber Defense Benchmark, Kimi K3 hit 33% coverage and DeepSeek 4 Flash hit 29.7% at a median cost of $0.17 per investigation, trailing only the open-weight leader GLM 5.2.
Source details: startupfortune.com
Why it matters
This development undermines the argument that certain AI capabilities are too dangerous for public release, as an accessible open- model is already demonstrating superior performance in the exact task cited as a risk. It highlights a shift in the AI security landscape where open models are becoming viable, cost-effective tools for vulnerability detection, potentially accelerating the adoption of AI in cybersecurity workflows without the need for exclusive vendor partnerships.
The core issue is the validity of gating AI models based on security risks. Google argued that Gemini 4 Argon's ability to find exploitable flaws was too risky for open release. However, the fact that an open- model like DeepSeek V4 Pro is already outperforming it in public tests with a published methodology undermines this justification.
This shift has practical implications for security teams. They can now use open- models on their own hardware without vendor gates, potentially reducing costs and increasing flexibility. The lower cost of running open-weight models for vulnerability detection makes them more accessible to a wider range of organizations.
The rapid advancement of open- models in security tasks suggests that the gap between open and closed models is narrowing. This could lead to a broader adoption of AI in cybersecurity, with open models playing a significant role in vulnerability detection and remediation.
The two-week gating window for Gemini 4 Argon suggests that Google itself did not view the risk as severe enough to hold the model indefinitely. This indicates a pragmatic approach to AI release, balancing security concerns with competitive pressures and the rapid pace of AI development.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
Monitor whether other open- models continue to close the gap with closed frontier models in security tasks. Watch for regulatory responses to the rapid capability growth of open-weight models in sensitive domains like cybersecurity. Observe if Google or other providers adjust their release strategies in response to the competitive pressure from open-weight alternatives.
Future benchmarks and real-world deployments of open- models in cybersecurity will be crucial to assess their practical impact. If open models continue to perform well in security tasks, they may become the standard for vulnerability detection.
Regulatory bodies may need to reconsider their approaches to , particularly regarding the release of models with capabilities that could be used for both defensive and offensive purposes. The success of open- models in security tasks may influence policy discussions.
The competitive dynamics between open and closed model providers will likely intensify. Closed model providers may need to differentiate their offerings through other means, such as integration, support, or specialized features, rather than relying on exclusive access to advanced capabilities.