AWS Ships Continuum for Code Vulnerabilities

AWS has released Continuum for code vulnerabilities into general availability in two regions, US East (N. Virginia) and US West (Oregon). The launch lands roughly four months after the service was announced in a gated preview. Continuum finds and validates vulnerabilities in application code and then produces fixes, either as patches suitable for automation or as pull requests opened against the connected repositories. Esra Kayabali, a principal solutions architect at AWS, is the person describing the mechanics in the company's materials.
The underlying idea is to build a context graph from the customer's repositories together with the surrounding AWS environment. An optional collector can also supply resource and network configuration from wherever the code actually runs, which gives the service more to reason about than source files alone.
AWS executives framed the launch around the volume of alerts rather than the difficulty of detection. Swami Sivasubramanian, VP of Agentic AI, wrote that most organizations running large application estates see thousands of findings a day. Christopher Rae, who leads AI security go-to-market at AWS, put it more directly, saying scanners surface thousands daily and that most of it becomes noise that does not matter in a given environment. That framing explains the design center: the scarce resource is not vulnerability discovery but the human judgment needed to triage it.
Continuum's answer is validation in place of severity ranking. Instead of ordering findings by a severity label, the service checks which issues are actually reachable by executing the exploit inside an isolated sandbox, then attaches proof-of-concept code and reproduction steps to the finding. Each result carries an evidence level, one example being a verdict proven through static analysis of the vulnerable code path. The output is an attack path plus what AWS calls a blast radius assessment, rather than a risk score on its own.
For developers, that distinction matters. A severity score is a guess about impact based on the class of bug; a sandbox reproduction is a demonstration that the bug is real and triggerable in the code as written. The trade-off is that execution-based proof only covers what can be run, and sandbox conditions are not the same as production conditions. This suggests teams should treat a proven finding as high-confidence and treat the absence of proof as unresolved rather than safe.
Pricing, which the launch announcement does not mention, has two dimensions on the published page. A context fee covers lines of code protected per month, starting at $200 per million lines and stepping down to $100 above ten billion lines. An inference fee charges $10 per Continuum inference unit, consumed as findings are processed. AWS's worked examples show how the two combine: one million lines with 100 vulnerabilities at one unit each totals $1,200, while 100 million lines with 10,000 vulnerabilities at two units each reaches $218,200. In the larger case, context accounts for $18,200 and inference for $200,000, or roughly $20 per validated finding.
That structure ties cost directly to the volume problem the executives described. The same thousands of daily findings that motivate the product also drive the inference bill, which means tuning what gets submitted for validation is a budget decision, not just a workflow one. In practice, teams on large codebases will want to understand what triggers an inference unit before rolling the service across every repository.
Responding on LinkedIn, Bernardo Meireles Correa, an engineer working on non-human and agent identity at Avatier, welcomed the improved ranking and raised two concerns. One is that a proof in an isolated sandbox establishes the code is exploitable but not that a path to it exists in the customer's actual environment. The other is that findings the service cannot prove still need a destination, or they quietly accumulate into what he called the safe pile.
The first concern is partly addressed by the optional collector, which gathers AWS resource and network configuration to judge reachability, though it is additional setup rather than a default behavior. The second has no stated answer in the AWS documentation, which describes handling for findings that can be proven.
On autonomy, the current posture is conservative. Every remediation begins in a mode where Continuum proposes a fix and a person approves it before it ships, and the console itself warns that a generated patch does not mean the issue is fixed. AWS says it plans to let Continuum apply certain classes of fixes without approval once confidence criteria, such as test coverage and historical success rate, are met.
Jorg Huser, a CTO and former AWS principal, replied to Rae questioning how that threshold should be set. He argued that moving from approving each merge to approving a class of fixes deserves its own acceptance test, and warned that an overloaded review queue could itself become the reason to grant more autonomy. He asked what would qualify a fix class for autonomous merging and what would send it back to human review, and said he would want evidence that a fix removes the vulnerability while preserving expected application behavior.
Google's CodeMender occupies adjacent ground and, according to the reporting, ships something close to that evidence: it proves exploitability by running an exploit, declines to apply patches itself, and validates each patch with an LLM-as-a-judge framework testing for program equivalence, so the fix addresses the root cause without altering intended behavior. That is the verification step the Continuum console caveat concedes the service does not perform.
The product boundary reported in July remains visible in the documentation. The Continuum user guide covers code vulnerabilities and directs readers to a separately titled AWS Security Agent User Guide for penetration testing and threat modeling. Only the code vulnerabilities capability reached general availability. Penetration testing is priced separately at $50 per task-hour, and threat modeling, announced in preview in June, has no published pricing. The earlier confusion, voiced by AWS serverless hero Yan Cui about two product names with overlapping capabilities and no clarity on which to use, appears to persist at the documentation level.
Why it matters: security and platform teams drowning in scanner output now have an option that trades alert volume for execution-backed proof, at a price that scales with both codebase size and how many findings they choose to validate, which makes triage policy a cost lever. For developers, the near-term change is pull requests with reproduction evidence arriving in repositories rather than another dashboard, and the approval step remains human. The open question, as practitioners have framed it, is what evidence standard would justify letting fixes merge on their own; that standard is the thing to watch before granting broader autonomy, and it is my inference that teams will want their own acceptance criteria defined before AWS enables unattended merging.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
AWS has made Continuum for code vulnerabilities generally available in two regions. The service proves exploitability by running real exploits in a sandbox, then opens fixes as pull requests. Published pricing charges per finding processed as well as per line of code, and practitioners are asking what would justify letting it merge fixes on its own. By Steef-Jan Wiggers