By ByteBot
AWS launched Continuum for code vulnerabilities into general availability on October 22. The headline claim: an 89% end-to-end fix rate on the CyberGym-E2E benchmark, more than 20 points above the previous best. As of this writing, no independent lab has reproduced that number. That gap — between the vendor claim and the verification reality — is the first thing you should understand before connecting your repositories.
What Continuum Actually Does
Most security scanners hand you a findings list and walk away. Continuum runs a five-stage pipeline that ends with a pull request, not a spreadsheet.
It starts by building a context graph from your AWS environment — mapping network topology, IAM relationships, and resource dependencies. This infrastructure awareness is what lets Continuum estimate business impact rather than assign a CVSS score. From there, it generates threat models and deploys discovery agents to surface candidate vulnerabilities in the code.
The part that sets Continuum apart is stage three: sandbox exploitability confirmation. Instead of flagging something as “potentially exploitable,” Continuum attempts to reproduce the vulnerability in an isolated environment and provides proof-of-concept code with reproduction steps. If it cannot reproduce the exploit, it says so. That distinction matters: your engineering team stops wasting cycles on false positives.
Stage four produces a blast radius assessment — a documented attack path and potential reach estimate. Stage five delivers a pull request with the fix, the exploitability proof, and the blast radius attached. The fix arrives with its own evidence brief. The full announcement is on the AWS blog.
Learn Mode First, Enforce Mode Later
Continuum ships with two operating modes. Learn mode is the starting point: the agent proposes fixes and your team reviews and approves each one. This is the right posture for onboarding — you get to see how Continuum handles your specific codebase before trusting it to act alone.
Enforce mode lets the agent act autonomously within guardrails you define. AWS says you activate it after “sufficient precision trust is established” — which they leave deliberately undefined. The honest interpretation: run Learn mode long enough to form your own opinion, then decide. This trust-gating model is becoming standard for agentic tools, and for anything touching production security it is the right call.
How It Compares
GitHub Copilot Autofix and Snyk Agent Fix are the closest comparables.
Copilot Autofix is tightly integrated into GitHub pull requests and powered by CodeQL. It does not verify exploitability in a sandbox, does not produce blast radius documentation, and has documented limitations with multi-file fixes. It works well if you are GitHub-native and your vulnerabilities are single-file CodeQL findings — which many are not.
Snyk Agent Fix supports GitHub, GitLab, Bitbucket, and Azure DevOps, making it the multi-platform option. But it explicitly documents single-file fix scope and has no sandbox verification. Snyk claims 80% fix accuracy — also vendor-reported.
Continuum’s technical edge is real: sandbox exploit verification and infrastructure-aware blast radius documentation are features neither competitor offers today. The constraints are equally real: it requires an AWS account, and at launch it operates only in US East (N. Virginia) and US West (Oregon).
The Pricing Reality
Continuum is not free. The context fee starts at $200 per million protected lines of code per month, with the rate declining at higher volumes. On top of that, there is a usage-based $10 per inference unit charge for each fix generation run. A 500,000-line codebase runs $100 per month in context fees before inference. At enterprise scale the tiered pricing becomes more competitive, but for mid-size teams this deserves an explicit budget conversation before you enable it.
About That 89% Number
CyberGym-E2E is a CTF-style benchmark designed to test AI agents on security tasks in controlled environments. The 89% result (819 of 920 tasks) is well above the previous best of 65.9%, which is legitimately impressive if it holds in independent testing. The caveat is that CTF benchmark performance and production codebase performance measure different things. Production codebases have idiosyncratic patterns, legacy debt, and multi-system data flows that controlled benchmarks rarely replicate.
AWS has not done anything wrong by reporting this — vendor benchmark disclosure is standard practice. The absence of independent verification reflects how recently Continuum went GA, not a sign of manipulation. Expect third-party analysis within weeks as security researchers get access.
What to Do Now
If you are on AWS, enable Continuum in Learn mode on a non-production repository and watch how it handles your codebase for a few weeks. The PR-based workflow means there is no downside risk — nothing merges without your review. Hold off on Enforce mode until independent benchmark results and your own Learn-mode data give you confidence. Service details and pricing are on the Continuum product page. At $200 per million lines, this is priced for teams that have already calculated what a security engineer’s time costs per finding.
