AWS Launches Continuum AI Code Security With Human-Reviewed Fixes and Benchmark Caveats
3 sources · Updated 28d ago · First seen Oct 23, 2026

AWS has made Continuum generally available to identify, validate and propose fixes for code vulnerabilities, initially in the US East (N. Virginia) and US West (Oregon) regions. The AI-powered service combines repository analysis with customers’ AWS infrastructure context to assess exploitability and prioritize findings by potential business impact. Supporting evidence can include static analysis or, when reproducible, sandboxed exploits with proof-of-concept code and reproduction steps. Continuum opens GitHub pull requests containing proposed changes, risk assessments and evidence; human review and implementation remain required.
AWS reported that Continuum completed 819 of 920 CyberGym-E2E tasks within 90 minutes per task, achieving an 89.0% success rate against a previous public high of 65.9%. That figure does not mean it fixed 89% of the benchmark’s targeted historical vulnerabilities: patches addressed those specific flaws in only 37.8% of tasks, because agents were not told which historical vulnerability to fix and could succeed by repairing other real flaws. The benchmark covers only C and C++ memory-safety vulnerabilities, and The New Stack reported no independent verification of AWS’s results. Security teams should validate effectiveness against their own codebases and retain human approval rather than treating the headline score as a general remediation rate.


