AWS lets its AI agent generate evidence for security flaws

AWS has made a small product change that says a lot about where security operations are heading.

AWS Security Agent can now generate verification scripts for penetration-test findings. A finding no longer needs to remain a written report with screenshots and reproduction steps. The agent can produce an executable script that the security team can use to check whether the vulnerability is actually present in its own environment.

This is not a broad consumer AI launch. For CISOs, platform teams and boards that care about software risk, it matters. It moves the AI agent closer to work that used to sit with experienced penetration testers: turning a technical finding into something reproducible, documented and possible to prioritize in a patch queue.

AWS says Security Agent now creates ready-to-run scripts for confirmed findings. Teams download the script, configure environment variables and run it against the target system to verify the vulnerability. The scripts include setup instructions, documented environment variables and redacted sensitive values.

AWS also adds the important warning in its documentation: the verification scripts are generated with generative AI. They should be reviewed before execution, and they should only be run against systems the team is authorized to test. That is not legal boilerplate. It is the core governance issue.

When AI moves from explaining vulnerabilities to generating code that can reproduce them, the risk changes. The script can be useful evidence in triage. It can also become a dangerous artefact if it is stored in the wrong place, shared too widely or run against the wrong environment. For companies that already struggle with who is allowed to run scanners, pentest tooling and proof-of-concept code, this becomes another control problem.

From report to evidence chain

Traditional security findings often get stuck in a familiar gap. The tester describes the weakness. The development team cannot reproduce it. Operations is unsure whether it applies to production, staging or an old service. Risk owners lack the evidence needed to prioritize.

AI-generated verification scripts can shorten that loop. They make the finding more concrete: set up the environment like this, run the test like this, expect this result. That gives teams a better basis for deciding whether a finding should be patched immediately, moved into a sprint or rejected.

But it also makes the security process more code-driven. A finding is no longer only an observation. It becomes a program. That means it should be handled as code: versioning, review, access control, logging and deletion.

That is the leadership point. The news is not that AWS added another console feature. The point is that AI agents are starting to produce operational security artefacts. They create material that can influence patch priority, risk assessment and change windows.

What leaders should decide

For enterprises using AWS, or considering agent-based security testing more broadly, the control questions are practical.

Who is allowed to generate and download these scripts? Should they be stored in a security platform, a developer repository or a ticketing system? Must they run from isolated environments? Who approves them before use against production-adjacent systems? How long should they be retained after the vulnerability is closed?

The second question is vendor governance. When a cloud provider or security tool generates reproducible test code, the company needs to understand what is logged, who can access the findings and how sensitive data is redacted. AWS says sensitive values are redacted in the scripts. That is useful, but it does not remove the need for internal controls.

The third consequence is patch velocity. If an AI agent can make more findings reproducible in less time, the bottleneck may move away from the security team. It may land with application owners, change boards and suppliers that cannot respond quickly enough. The question is no longer only whether the vulnerability was found. It is whether the organization can close it.

This is a narrow AWS update with a wider operating lesson: AI in security is less about prettier reports and more about evidence chains, permissions and speed. The companies that benefit most will not be the ones that let agents run freely. They will be the ones that define where scripts can be created, who can use them and how the results feed into patch management.

Sources and media

• Primary source: AWS, “AWS Security Agent adds verification scripts for pentest findings”, published May 22, 2026: https://aws.amazon.com/about-aws/whats-new/2026/05/aws-security-agent/
• AWS Security Agent product information: https://aws.amazon.com/security-agent/
• AWS documentation on reviewing findings and verification scripts: https://docs.aws.amazon.com/securityagent/latest/userguide/review-penetration-findings.html
• Thumbnail: OpenAI Image 2 / hogby.ai.