Model Watch High For CISO, SOC / IR, AppSec, Red Team AI-assisted
Vals AI: open-weight DeepSeek V4.1 Flash nears closed model on cyber-range tests
Vals AI's CyberOffense benchmark ran AI agents solo against simulated corporate networks, starting from a public web address. OpenAI's restricted GPT-6 Sol Daybreak scored best at 80.8 percent of objectives, and the open-weight DeepSeek V4.1 Flash 1017 scored 72.8 percent at about $3 per run versus about $22. UK AISI reports the open-vs-closed cyber capability gap has shrunk to four to seven months.
Why it matters
Capable offensive agents are increasingly available from open-weight models that no provider can restrict, so organizations cannot rely on closed-model access controls alone. Attackers gain low-cost automation for breaching flat or weakly segmented corporate networks.
Do this
- Review internal network segmentation and credential reuse, since the tested agents chained web access to Windows admin privileges and reused stolen passwords.
- Run AI-assisted vulnerability discovery and remediation on your own externally facing applications and internal hosts, as Vals suggests defenders can use the same model capabilities.
- Hunt for unusual high-volume, automated sequences of reconnaissance, login attempts, and lateral movement from single hosts, and alert on them.
- Enforce unique, rotated privileged passwords and limit admin session reuse across servers to reduce the value of stolen credentials.
- Reassess threat models for long-running agent activity, not only short tests, and track AISI and Vals updates on open-model cyber capability.
Signals
- NoCISA KEV
- 30/100Priority
- 60%Confidence
Evidence
“Deepseek V4.1 Flash 1017, a freely available open-weight model, isn't far behind at 72.8 percent and costs only about $3 per run instead of roughly $22.”
“Once a model's weights are public, no provider can control or restrict how it's used, Vals researchers say.”
“The gap between open and closed frontier models in cyber capabilities has shrunk to four to seven months.”
Sources
Affected / related
GPT-6 Sol Daybreak, DeepSeek V4.1 Flash 1017, Vals CyberOffense benchmark, Vals AI, UK AI Security Institute
Published 2026-11-07 15:35 UTC · updated 2026-11-08 04:27 UTC · AI-assisted summary, checked against quoted sources · How we score