Wire · technology
AI struggles to patch vulns without adult supervision
◆ Sectors
AI Frontier ModelsCybersecurity
◆ Source
◆ Verified
Fusion42 · 30 August 2026 · Fusion42 review
Research from 1Password's Off-by-1 Labs shows that AI models like ChatGPT 5.5 and Claude Opus 4.8 autonomously create fully effective security patches only about 26% of the time, often requiring human oversight to avoid failures or new vulnerabilities.
This Wire brief sits within Fusion42's coverage of AI Frontier Models and Cybersecurity.
◆ ◆ The Wire takeaway
You cannot fully trust AI to fix your software vulnerabilities without human review; this means your security team needs to keep control and scrutinise AI patches carefully to avoid risks and rework.
◆ Coverage
1 source · 6 Aug 2026
◆ Related on Wire
◆ Topics
AI Frontier ModelsCybersecurityai-patchingsecurity-vulnerabilitiesllm-limitationshuman-reviewcybersecurity