AI agents breach enterprise networks, and the gap narrows
Vals AI shows that AI agents can independently penetrate simulated enterprise networks. Open models quickly catch up with closed models, at much lower costs, according to several recent studies.
AI agents are now capable of independently attacking simulated enterprise networks without human assistance, according to the CyberOffense benchmark by Vals AI published on November 7, 2026. The test pits models against a public web address, then measures their progress toward administrator privileges in a realistic environment. The gap between open and closed models is narrowing, changing the game for both cybersecurity defenders and attackers.
When a model learns to force a door, defense must keep up
CyberOffense is based on a simple and rather harsh logic: an agent starts from an isolated Kali Linux machine, explores a simulated network, executes commands in sequence, then attempts to escalate to admin rights. Each action is logged and automatically scored. GPT-6 Sol, used in OpenAI’s restricted cyber program Daybreak, achieves 80.8% of objectives, while Deepseek V4.1 Flash 1017, a freely available open-weight model, reaches 72.8% at a cost of about 3 dollars per run, compared to about 22 dollars for GPT-6 Sol. In other words, automated attacks are no longer very expensive. This is no laughing matter for CISOs.
The cost drops, capability rises
In the more difficult Lighthouse test, Daybreak takes admin control of the first fictional company, steals passwords, then uses them to infiltrate a second company. Vals reports 27 objectives achieved out of 31 in under three hours. Deepseek completes 21 at a much lower cost, without being limited only by security guardrails. Most importantly, execution time matters greatly: the open-weight model achieves no results in its first 25 million tokens, then makes significant progress after 100 million. Too short benchmarks therefore underestimate the real capability of some open models.
Open and closed: only four to seven months apart?
Vals’ results align with an analysis by the UK’s AI Security Institute published in July 2026. According to this study, the lag between leading open and closed models in cyber capabilities has dropped to four to seven months, compared to six to ten months for much of 2025. The decrease is clear. It is all the more significant since open models remain much cheaper to run: the AISI estimates a cyber-range run at 1.19 dollars for Deepseek V4-Pro, versus 85 dollars for Anthropic Opus 4.5. Meanwhile, security protections tested on open models prove largely ineffective.
What this means for defenders
Vals also sees an opportunity for companies: the same models capable of exploiting a vulnerability can be used to find and fix it upstream. But the underlying signal remains less reassuring. Attacks on real networks are expected to increase in the coming months, according to Vals AI, especially as other research points in the same direction. An Anthropic study showed that the open model GLM-5-3 can build exploits at a level close to its most advanced internal model, Mythos. The market no longer faces just lab demonstrations; it is looking at a tool that is settling in, costing less, and advancing rapidly.
Key points
- Vals AI’s CyberOffense was published on November 7, 2026.
- GPT-6 Sol achieves 80.8% of objectives on CyberOffense.
- Deepseek V4.1 Flash 1017 reaches 72.8% at lower cost.
- AISI reports a four to seven month gap in July 2026.
- Vals predicts more attacks on real networks.
- Anthropic observed GLM-5-3 close to Mythos.
By the numbers
- 80.8% — objectives reached by GPT-6 Sol on CyberOffense (Vals AI, 2026).
- 72.8% — objectives reached by Deepseek V4.1 Flash 1017 (Vals AI, 2026).
- 4 to 7 months — estimated gap between open and closed models in cyber (AISI, July 2026).
- 1.19 dollars — cost of a Deepseek V4-Pro cyber-range run (AISI, 2026).
- 85 dollars — cost of a comparable run with Anthropic Opus 4.5 (AISI, 2026).