AWS generalizes its vulnerability fixing agent, with a score to be taken with caution

AWS makes Continuum, its security agent for code, generally available after scoring 89% on CyberGym-E2E. The figure is impressive, but the test also highlights the limits of benchmarks in software security.

AWS made Continuum generally available on October 24, 2026, a security agent that detects vulnerabilities in code, validates them in the repository, then proposes fixes as pull requests. The announcement matters for SRE and security teams as it promises to automate part of the hunting and remediation work. But the headline score, 89% on CyberGym-E2E, should be interpreted with caution.

Un agent qui cherche, vérifie et propose des correctifs

According to AWS, Continuum analyzes repositories, cross-checks detected vulnerabilities with the client’s AWS environment, then opens a merge request with a suggested fix. The service is therefore not a simple static scanner: it attempts to reproduce the problem, then fix it without breaking existing tests. This is where the promise becomes interesting, and a bit less glamorous than a demo press release. Software security loves numbers, but code prefers to be fixed.

AWS also presents Continuum as a tool designed for real production scenarios. The provider says it respected the benchmark rules, including internet access limits, and verified that the agent did not just find already public fixes. In this context, the challenge is not only to detect bugs but to handle them at the scale of a real software fleet.

Le 89% qui fait parler, et ce qu’il ne dit pas

The highlighted figure comes from CyberGym-E2E, where AWS claims Continuum solved 819 tasks out of 920, i.e., an 89.0% end-to-end success rate. The company adds that a public competitor had previously reached 65.9%. On the same test bench, the agent found and reproduced a bug in 92.5% of cases, stopped the crash in 89.6% of cases, and maintained project tests in 89.0% of cases. However, it only fixed the historical vulnerability targeted by the benchmark in 37.8% of tasks.

This last point changes the interpretation. CyberGym-E2E provides no flaw description, proof of concept, crash log, or existing patch; the agent has 90 minutes to find and fix a problem in one of the 139 open source projects tested. The overall score thus shows that an agent can often produce a useful fix, but not necessarily the one the benchmark had in mind. AWS also acknowledges that extending the runtime raises the result to 93.7% [to be verified according to the exact published test conditions].

Ce que cela change pour les équipes de sécurité

The general availability of Continuum comes as vulnerabilities find their fixers faster… and sometimes their discoverers too, since AI accelerates vulnerability research. For security managers, the benefit is clear: reduce the delay between detection, reproduction, and fix proposal. However, automation does not solve everything. It remains to be seen how the tool behaves when the team asks it to fix a specific flaw, in a complex repository under business constraints.

The real issue is therefore not the 89% itself, but the shift from a tool that alerts to a tool that proposes an actionable remediation. If AWS delivers on this promise across varied environments, teams’ work could evolve toward more validation and less manual triage. Otherwise, Continuum will remain a good seller of statistics. And a bit of humility never hurts a benchmark.

Key points

  • Continuum became generally available on October 24, 2026.
  • AWS announces an 89.0% success rate on CyberGym-E2E.
  • The benchmark covers 139 open source projects in 90 minutes.
  • The score drops to 37.8% on the targeted historical vulnerability.
  • Matt Garman calls it a "genuinely transformative for defenders" tool.
  • AWS says 93.7% is achievable with more time.

By the numbers

  • 920 tasks — total tested on CyberGym-E2E according to AWS, October 2026.
  • 819 tasks — resolved by Continuum, i.e., 89.0%, according to AWS.
  • 139 open source projects — scope of the CyberGym-E2E benchmark.
  • 90 minutes — time allotted to the agent per task.
  • 37.8% — resolution rate of the targeted historical vulnerability.

To read