We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·18d ago·5 min read
How Far Behind the Frontier are Leading Open Weight Models on Cyber?
.png&w=3840&q=75)
TL;DR
Recent evaluations show that leading open weight AI models, GLM-5.2 and DeepSeek V4-Pro, are closing the cyber capability gap with closed models. The gap has narrowed from 6-10 months to just 4-7 months.
✦ Why It Matters
Cybersecurity teams should prioritize monitoring advancements in open weight models to anticipate potential risks and vulnerabilities.
Key Takeaways
- Evaluate your current cybersecurity measures against emerging AI threats.
- Consider investing in closed weight models for enhanced safety.
- Monitor advancements in open weight models for potential risks.
- Collaborate with cybersecurity experts to strengthen defenses.
Full Summary
Since 2023, AISI has tracked the cyber capabilities of AI models, finding that closed weight models consistently outperform open weight models. Open weight models, such as GLM-5.2 and DeepSeek V4-Pro, offer benefits like adaptability and cost-effectiveness but lack the safety measures of their closed counterparts.
Evaluations showed GLM-5.2 performing comparably to models released 4 to 7 months earlier on narrow cyber tasks, indicating a reduced gap from previous assessments. However, on more complex cyber ranges, the gap remains larger, suggesting limitations in sustained autonomous cyberattack execution.
As frontier AI capabilities advance, the findings emphasize the need for proactive measures in cybersecurity to mitigate risks associated with open weight models.