AWS Put the Proof Before the Patch
The interesting part of AWS Continuum is not that an AI agent found vulnerabilities.
It is that AWS is packaging the evidence loop around the finding: reproduce the issue, test the repair, preserve the attack path, and keep a human in the approval path before anything ships.
That is a more important product boundary than the model behind the agent.
What changed
AWS announced general availability for Continuum for code vulnerabilities on October 22, 2026. AWS describes it as a service that finds and validates code vulnerabilities, ranks them using business context, and delivers fixes as patches or pull requests.
The service is available in US East (N. Virginia) and US West (Oregon). AWS says every remediation starts in a mode where Continuum proposes a fix and a person approves it before it ships.
The launch followed an AWS Security Blog report on a CyberGym-E2E evaluation. AWS says Continuum passed 819 of 920 tasks within the benchmark’s 90-minute limit, for an 89.0% end-to-end success rate. That is 23.1 percentage points above the previous public high AWS cites.
Those numbers are AWS’s own evaluation results. They are useful evidence about the system AWS built, not an independent replication or a guarantee of production performance.
The benchmark measures a loop, not a clever answer
CyberGym-E2E is designed around the full vulnerability lifecycle. An agent receives a vulnerable revision of a real open source project and must find a vulnerability, demonstrate it with an input, produce a patch, and preserve the project’s tested functionality.
The benchmark contains 920 tasks across 139 open source projects. AWS says the median project contains more than 600,000 lines of code. External network access is blocked, and protected benchmark files cannot be modified.
The stages are cumulative:
| Stage | What it measures | Continuum’s reported result |
|---|---|---|
| S1 | Finds and reproduces a vulnerability | 92.5% |
| S2 | Repairs its generated crash | 89.6% |
| S3 | Preserves tested functionality | 89.0% |
| S4 | Repairs the benchmark’s selected historical vulnerability | 37.8% |
S3 is the main end-to-end measure in the AWS report. S4 is diagnostic because a repository can contain several valid vulnerabilities, so a system can repair a real flaw without repairing the benchmark’s selected target.
That distinction matters. A security agent should not be judged only by whether it can point at suspicious code. It should be judged by whether it can produce evidence that survives an adversarial test and a regression suite.
The architecture is the story
AWS describes Continuum as a multi-agent system with specialized phases for discovery, validation, and remediation.
Discovery records candidate findings and the code evidence behind them. Validation attempts to turn a candidate into a demonstrated issue by constructing a proof of concept and running it against the vulnerable program. Remediation traces the root cause, produces a patch, and checks both that the demonstrated failure is blocked and that the project’s functionality tests still pass.
The same evidence moves through the phases. That gives the system a continuous record from suspicious code to exploit demonstration to tested repair.
This is the difference between an AI scanner and an AI security system. A scanner emits findings. A system has to establish what is real, what is reachable, what changed, and whether the change closed the path without breaking something else.
Context beats severity labels
Continuum also builds a context graph from code repositories and AWS environment data. AWS says the graph can capture resources, network configuration, ownership, dependencies, and the data touched by each service.
That lets it ask a more useful question than “what is the CVSS score?” A critical flaw in a development-only utility is not the same operational problem as a medium flaw one network hop from a payments database.
For a finding that can be reproduced, AWS says Continuum runs the exploit in an isolated sandbox and attaches proof-of-concept code, reproduction steps, source analysis, and a blast-radius assessment. It also evaluates whether existing defenses would detect or stop the validated path.
This is a better shape for automation because the output is not just a number. It is a claim with supporting evidence and an explicit boundary around impact.
The approval gate is not a temporary detail
AWS says Continuum opens confirmed fixes as pull requests. The generated patch does not mean the issue is fixed. A team still reviews and merges it.
AWS also describes a future mode in which teams can automate selected remediation classes once confidence gates pass, such as test coverage, scope bounds, and historical success rate.
That progression is sensible, but it changes the meaning of “autonomous.” Autonomy should not be a boolean attached to the model. It should be a graduated permission set attached to evidence.
A system might be allowed to open a pull request automatically, but not merge it. It might be allowed to upgrade a dependency with strong test coverage, but not edit an authentication boundary. It might be allowed to apply a narrowly scoped configuration change only when the exploit reproduction, rollback path, and monitoring signal are all present.
The control loop has to record those distinctions. Otherwise an approval gate becomes a ceremonial button rather than a trust boundary.
What the benchmark does not measure
CyberGym-E2E is a meaningful step toward end-to-end evaluation, but it is not a complete production security test.
AWS says the benchmark focuses on memory-safety vulnerabilities in C and C++ projects because sanitizer-detected crashes provide objective evidence. That leaves many languages and vulnerability classes outside its current scope, including common production issues that do not reduce to a reproducible crash.
The benchmark also does not provide deployment context. A production decision depends on service exposure, permissions, network paths, data sensitivity, compensating controls, and the cost of a bad patch. AWS’s environment graph is intended to supply that context in the product, but it is outside the benchmark result itself.
The right conclusion is therefore narrower than “AI solved vulnerability management.” The reported result shows that a structured system can make measurable progress across discovery, validation, and repair under a fixed test contract.
What agent builders should copy
The useful design pattern is available even if you never run AWS Continuum:
- Give discovery, validation, and remediation separate responsibilities.
- Carry evidence forward instead of passing only a natural-language summary.
- Reproduce the claimed failure in an isolated environment.
- Test the repair against both the proof of concept and regression behavior.
- Add deployment context before ranking business risk.
- Bind actions to explicit approval and confidence gates.
- Keep the full trace, including the model, tools, patch, tests, and reviewer.
The agent that finds a bug should not be the only authority that decides the bug is real. The agent that writes a patch should not be the only authority that decides the patch is safe.
That separation is not overhead around the model. It is the security product.