Two AWS announcements, two different products
Start with the disagreement, because it decides what a buyer thinks they are getting.
The short What's New announcement is a tidy progression. Continuum “starts in Learn mode where it proposes the remediation fixes that you approve. When it builds sufficient trust in its precision, you switch to the Enforce mode and let it act autonomous within the guardrails you set.”
The AWS News Blog post, published the same day and written as a walkthrough of the actual console, puts the same step in a different tense, twice. “Every remediation starts in a mode where Continuum proposes the fix and a person approves it before it ships. In the future, you will be able to decide which actions Continuum can take on its own, within the guardrails you define.” Later, on the fix itself: “Every fix requires human approval: you review and merge the pull request, so the final decision stays with your team. In the future, you will be able to build confidence in specific remediation classes and let Continuum apply those fixes automatically once its confidence gates pass.”
Those are not two descriptions of one switch. One says the switch exists once trust is built. The other says the switch is coming. Neither announcement defines the precision threshold, the guardrails, or the confidence gates beyond naming examples such as test coverage, scope bounds and historical success rate. AWS's detailed GA walkthrough describes human-reviewed pull requests as the current workflow. Its shorter announcement appears to describe an autonomous Enforce mode, leaving that mode's GA availability unclear. Until AWS reconciles its own texts, that is where the question stays: the workflow AWS demonstrates in detail is review-first, and the autonomy is announced in two incompatible tenses.
The part that is actually new
Strip the autonomy question away and the remaining product is still a real change in shape. Scanner output is normally a list ordered by a generic severity score, which is why a critical rating on a development utility can sit above a medium rating on a service one step from a payments database. Continuum's pitch is to replace that ordering with evidence: a context graph built from your repositories and your AWS environment, a finding tested in an isolated sandbox, and an issue record that carries the attack path, the supporting source analysis and a blast-radius assessment instead of a score alone.
That evidence travels with the fix. The pull request description carries the same risk assessment, remediation summary and proof as the console, so the engineer reviewing it does not have to take the tool's word for why this patch, in this order. In AWS's own walkthrough demonstration, one demo issue in a cart service produced a pull request changing four files, with 50 lines added and 24 removed — a vendor demonstration, not a customer result. AWS is also explicit, in the console guidance its walkthrough quotes, that a generated patch does not mean the issue is fixed: the reviewer still merges, and the final decision stays with the team.
The two GA announcements provide no new, product-wide precision, false-positive or remediation-success benchmark, and neither publishes a list of supported languages. AWS has published separate security-agent testing — an October evaluation of its penetration-testing agent, now part of Continuum — but those results should not be treated as a comprehensive independent evaluation of Continuum's GA code-vulnerability workflow. AWS says the service selects among multiple frontier cyber models for each task, and does not name them. The evaluation a GA buyer can inspect, for now, is the walkthrough.
The bill, in AWS's own arithmetic
Pricing is at least public, and its structure tells you where the cost risk sits. The context fee is predictable: a monthly charge per million protected lines of code, tiered — $200 per million for the first 10 million lines, $180 for the next 90 million, $160 above 100 million through 1 billion, $130 above 1 billion through 10 billion, and $100 above 10 billion. The inference fee is the variable part, at $10 per Continuum inference unit, where a unit is a quantity AWS defines and the customer does not directly control. AWS says Continuum “selects and blends the right models for each task behind the scenes, so you see a single unit rather than individual model charges” — convenient, and opaque in the same sentence.
AWS's worked examples are illustrative, not bills, and each carries its assumption on its face. A small codebase — 1 million lines and 100 vulnerabilities in a month — comes to an estimated $1,200, on the stated assumption of an average of 1 inference unit per vulnerability. A mid-size example, 10 million lines and 1,000 vulnerabilities at the same assumed average, comes to $12,000. The enterprise example — 100 million lines and 10,000 vulnerabilities at an assumed average of 2 units each — comes to $218,200, of which $18,200 is context fees and $200,000 is inference. Change the assumed units per vulnerability and the total moves with it; AWS notes actual inference varies with the work each vulnerability requires. The published code-vulnerability pricing examples do not specify a maximum monthly inference charge or an automatic spending cap. Buyers should confirm available budget controls and billing alerts with AWS.
So the buying question is not the context fee. It is how many inference units a month of your findings actually consumes, and that number is only knowable by running it. Treat the first month as a metering exercise, in the two regions where it exists, and budget on the workflow AWS's walkthrough demonstrates — human-reviewed pull requests — until AWS clarifies what Enforce mode includes at GA.
Sources
- [1] AWS News Blog, “AWS Continuum for code vulnerabilities is now generally available”, AWS, 22 October 2026 — primary source, read in full first-hand for this article: the GA announcement; the context-graph and sandbox validation description; the walkthrough demonstration in which a cart-service issue produced a pull request changing four files, with 50 lines added and 24 removed; the console guidance that a generated patch does not mean the issue is fixed; both future-tense autonomy passages (“Every remediation starts in a mode where Continuum proposes the fix and a person approves it before it ships. In the future, you will be able to decide which actions Continuum can take on its own…”) and the confidence-gate examples (test coverage, scope bounds, historical success rate); and the two-region availability.Read source
- [2] AWS What's New, “AWS Continuum for code vulnerabilities is now generally available”, AWS, 22 October 2026, timestamped 2026-10-22T21:00:00Z — the short announcement carrying the present-tense Learn mode / Enforce mode passage (“starts in Learn mode where it proposes the remediation fixes that you approve. When it builds sufficient trust in its precision, you switch to the Enforce mode and let it act autonomous within the guardrails you set”), the “picks the right AI model for each task” wording, and the GuardDuty / Security Hub line. Read for this article via the announcement's full-text mirror on GIXtools; the same Learn/Enforce wording is independently reflected in The AWS News Feed's republication of the What's New item and in OOSSA's report on the announcement.Read source
- [3] AWS Continuum pricing page, AWS — read in full first-hand on 23 October 2026: the code-vulnerabilities section's context-fee tier table ($200 per million lines for the first 10 million, $180 for the next 90 million, $160 above 100 million through 1 billion, $130 above 1 billion through 10 billion, $100 above 10 billion), the $10 per CIU inference fee, the “selects and blends the right models for each task behind the scenes” wording, and all three worked examples with their stated CIU-per-vulnerability assumptions. The same page separately prices Continuum for penetration testing, a different capability; its rates are not used anywhere in this article.Read source
- [4] AWS Security Blog, “AWS Security Agent (now part of AWS Continuum) exploited 33 of 37 new real-world CVEs”, AWS, 15 October 2026 — read in full first-hand: a separate evaluation of the autonomous penetration-testing agent (now part of Continuum), run on Claude Opus 5 against 37 real, recently disclosed CVEs packaged as live containerized targets; 35 flags were captured and 33 of 37 were captured through the intended vulnerability end-to-end, confirmed by a per-sample integrity audit. Cited only to scope this article's separate-testing sentence; it is not a benchmark of the GA code-vulnerabilities workflow, and no figure from it is used as one.Read source