Skip to content
← Back to the blog

Cyber Capabilities: Open Models Are Narrowing the Gap with Closed Ones

  • Governance e compliance
Cyber Capabilities: Open Models Are Narrowing the Gap with Closed Ones

A new Vals AI benchmark shows that the advantage of the most advanced closed models in offensive cyber capabilities remains real, but it is narrowing. The test, called CyberOffense, simulates realistic corporate networks and measures how far an AI agent can progress starting from a simple public web address, without human help. In the result reported by The Decoder, GPT-6 Sol Daybreak delivers the best performance, but an open-weight model such as DeepSeek V4.1 Flash comes very close, at a significantly lower cost per run.

The most relevant aspect is not only the comparison between scores, but also the type of scenario tested. The agents operate from an isolated Kali Linux machine, send commands to the simulated network, move through web applications and Windows systems, and aim for administrative privileges. In the most complex test, OpenAI’s model completes 27 of 31 objectives in less than three hours; DeepSeek completes 21, at an indicated cost of about one tenth. Vals also notes that short tests can underestimate the potential of open models, because some capabilities emerge only in very long runs and with high token consumption.

For those who govern security, risk and compliance, the key point is that the entry barrier to sophisticated offensive capabilities is lowering. If relatively inexpensive open-weight models achieve results close to those of closed systems, it becomes harder to rely on access limits, vendor controls or high costs as containment factors. The source also cites earlier analyses from the UK AI Security Institute and Anthropic, which point in the same direction: the time gap between the closed frontier and open models is shrinking, while the safety measures built into open models prove largely ineffective once the weights are made public.

This changes the defense timeline. For companies and public administrations, it is no longer enough to monitor only the major vendors or the best-known releases: it is necessary to assume that capabilities once restricted can spread quickly into environments that are harder to control. The same dynamic, Vals notes, can also be used defensively, by using these models to identify and fix vulnerabilities before attackers do. For veriax.ai’s audience, the news reinforces the need for operational model governance, with continuous risk assessments, usage controls and separate test environments, especially in regulated contexts where response times and traceability are already essential constraints.

Source: The Decoder, November 7, 2026