AWS Security Agent Exploited 33 of 37 Recent CVEs in Internal Test
AWS says its Security Agent successfully exploited 33 of 37 recently disclosed CVEs in an isolated test following an internal audit. Two attempts rendered the tested target inoperable.
AWS Security Agent successfully exploited 33 of 37 recently disclosed CVEs through the intended path in an internal AWS test, according to the company. AWS published the 89.2% result on October 15, 2026, and the agent is now part of the AWS Continuum platform.
The test focused on an end-to-end process: starting from a low-privilege or unauthenticated position, the agent had to find a way to exploit a specific flaw and obtain a marker confirming success. AWS says it used the Claude Opus 5 model and set a limit of ten hours for each sample.
AWS Security Agent Tested Recent CVEs
The sample included 37 CVEs from 2025 and 2026 across eight vulnerability classes. The targets were not production systems but isolated, containerized environments. These ran either a vulnerable version of upstream software or a faithful minimal reproduction of the issue.
At the start, the agent received a URL, a one-sentence hint, and unauthenticated or low-privilege access. AWS initially recorded 35 obtained markers. After review, however, it excluded two of those captures because the agent obtained the marker through a route other than the intended one. The company therefore reports a final result of 33 successes out of 37 attempts.
In two other cases, AWS says the agent reached the point of exploiting the vulnerability but then rendered the target inoperable before reading the marker. This detail shows that successfully finding an exploitation path does not necessarily mean that the entire test procedure was completed safely.
A Different Type of Evaluation Than CyberGym-E2E
The new result does not concern the same type of tasks as the earlier CyberGym-E2E benchmark associated with AWS Continuum. That benchmark evaluated procedures involving memory-safety flaws in code, while the current test tracks the reproduction and exploitation of recent, real-world CVEs in prepared environments.
The score therefore cannot be directly compared with the previous 89% result in CyberGym-E2E. The task set, evaluation target, and methodology differ. AWS also describes its benchmark as continuously expanded and methodologically designed to limit training-data contamination. This claim alone, however, does not prove complete resistance to memorizing publicly available exploits.
Functions and Usage Limits
AWS documentation places AWS Security Agent within AWS Continuum. Its listed capabilities include authorized penetration testing, design review, and GitHub integration.
The internal benchmark result does not, however, confirm the same success rate or safety when testing customers’ production systems. A complete public dataset, independent reproduction, or audit of the individual CVE trajectories has not yet been provided. The two cases in which the agent rendered the tested target inoperable are also a relevant limitation.
For autonomous tools capable of creating a functional exploit from a low-privilege starting point, precise definition of the authorized scope and human oversight are therefore important. According to its announcement, AWS is working on safeguards against destructive exploitation and on restoring the target’s state.
What Will Be Important to Watch
Further insight into the tool’s capabilities could come from technical materials enabling independent verification, results on a larger externally audited set of current CVEs, and details about planned safeguards. AWS has not yet provided information about the availability, pricing, regions, or limitations of the penetration-testing feature within AWS Continuum.
Sources
- AWS Security Blog – Confirms the announcement date, internal test methodology, the audited result of 33 out of 37, two excluded captures, two failures after rendering the target inoperable, and planned safeguards.
- AWS Documentation – Confirms that AWS Security Agent is documented as part of AWS Continuum and describes its functions, including penetration testing.
Verified and updated: October 16, 2026 06:20
