Home / Technology / AI finds flaws faster than they can be fixed
AI finds flaws faster than they can be fixed
10 Oct
Summary
- Anthropic's AI identified nearly 30,000 software findings in one year.
- Only 516 of 6,157 disclosed findings were patched upstream.
- New programs aim to shift more validation to providers and maintainers.
Anthropic's advanced AI models are uncovering software vulnerabilities at an unprecedented rate, identifying 29,439 findings between November 1, 2025, and October 2, 2026. Of the 6,157 findings disclosed to maintainers, only 516 were patched upstream, highlighting a significant gap between vulnerability discovery and remediation.
To combat this, Anthropic launched the Cyber Mission on October 8, 2026. This initiative, alongside the OSS Scanner program, aims to accelerate the patching process by involving providers and maintainers more directly in the validation and fixing of identified weaknesses.
The Critical Infrastructure Defense Program, part of the Cyber Mission, partners with 11 major companies to protect essential services. This program is an expansion of Anthropic's efforts, following a successful initiative for state, local, tribal, and territorial governments that reached over half of US states.
Challenges remain, particularly with critical infrastructure, where fixing vulnerabilities can take months or even decades due to the risks of disrupting operations. A proposed $100 million bill, the AI Cyber Defense Act, seeks to authorize funding for pilot programs to grant critical infrastructure operators access to frontier AI models for defense.
