Open-weight models now match frontier cyber performance from just four months ago at a fraction of the cost

2026-07-18 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

An analysis by the British AI Security Institute (AISI) reveals that open-weight AI models now closely trail closed frontier systems in cyber capabilities, with the gap shrinking from 6-10 months to 4-7 months. Models like GLM-5.2 (released June 2026) and DeepSeek V4-Pro matched the performance of older closed systems such as Opus 4.6 (February 2026) and Opus 4.5 (November 2025) on 70 "Narrow Cyber Tasks" covering vulnerability research and reverse engineering. In "Cyber Ranges" simulating a 32-step network attack, GLM-5.2 performed similarly to Opus 4.5, while GPT-5.6-Sol and Claude Mythos 5 nearly completed the full simulation. Open models are significantly cheaper, with a 100-million-token Cyber Range test costing \$46 for GLM-5.2 versus \$85 for Opus 4.5/4.6, and DeepSeek V4-Pro costing just \$1.19. However, their safeguards are easily bypassed, posing a "persistent and irreversible risk of misuse" and reducing the time cyber defenders have to prepare for new attack types.

Key takeaway

For AI Security Engineers assessing emerging threats, the rapid advancement and low cost of open-weight models like GLM-5.2 and DeepSeek V4-Pro mean your defensive strategies must adapt faster. These models offer powerful cyber capabilities with easily bypassed safeguards, significantly shortening the window for preparation. You should prioritize developing proactive defenses that do not rely on model-level controls, focusing instead on network-level resilience and rapid threat intelligence integration.

Key insights

Open-weight AI models now nearly match frontier cyber capabilities from months ago at a fraction of the cost, but with easily bypassed safeguards.

Principles

• Open models close cyber capability gaps rapidly.
• Cost-effectiveness scales cyberattack potential.
• Access control is key to AI safety measures.

Method

AISI tested models using "Narrow Cyber Tasks" (70 tasks across four difficulty levels) and "Cyber Ranges" (simulated 32-step network attacks on corporate networks).

In practice

• Evaluate open models like GLM-5.2 for cost-effective cyber tasks.
• Anticipate rapid evolution of open-source cyber tools.
• Prioritize robust, non-model-dependent cyber defenses.

Topics

• Open-weight AI Models
• Cyber Security
• AI Safety
• Large Language Models
• Threat Intelligence
• Autonomous Cyber Capabilities

Best for: CTO, VP of Engineering/Data, Research Scientist, AI Security Engineer, AI Scientist, Director of AI/ML

Related on AIssential

• GLM-5.2: The real security risk? Plus: Vibe hunting, the end of CVSS and updates on Lightwell — Open-weight AI models pose significant security risks by enabling advanced capabilities without traditional safeguards, necessitating proactive defense.
• [AINews] AI Cybersecurity becomes top of mind — AI models, especially cyber-capable ones, require adversarial containment and robust evaluation infrastructure to prevent unintended escapes and goal-directed reward hacking.
• Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan — The AI frontier is rapidly diffusing, challenging control mechanisms and necessitating new regulatory and security paradigms.
• not much happened today — AI capabilities are rapidly advancing in cyber defense, general computer agency, and cost-efficient open-weight models.
• How Open-Weight Models Changed the AI Landscape — Open-weight models drive rapid AI innovation through shared designs and iterative improvements.
• AI Just Entered A New Era — GLM 5.2 demonstrates open-weight AI can rapidly approach Frontier-level capabilities through innovative training and architectural advancements.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.