AWS launches on-demand pentesting driven by an AI agent
AWS is putting on-demand pentesting, driven by an AI agent, on the menu. The service is called AWS Continuum for penetration testing, formerly AWS Security Agent, renamed after joining the Continuum portfolio. You give it an application and access, it builds its own context, designs multi-step attack scenarios and tests whether vulnerabilities are actually exploitable, not just present on a scanner report.
How it works
The workflow described in AWS documentation has three phases: the agent discovers the application and its scope, drafts plausible attack chains, then executes them to verify real exploitability. The promise is clear: move past scanner reports listing theoretical CVEs and produce actual proof of exploitation.
Testing authenticated areas
Before the test, you can provide login credentials. The agent authenticates like a real user and discovers accessible domains on its own, then proposes them as the test scope. This solves a genuine pain point: authenticated areas are exactly what classic scanners miss.
The trade-off is well known: you hand your credentials to AWS. For production environments, the question of granted scope and agent action auditability matters even more than the results.
What we don't know yet
AWS pitches a "frontier" agent covering the whole development cycle. Two open questions remain. Billing is based on cumulative task hours, which makes the final cost hard to estimate upfront. And no independent benchmark on false positive rates has been published yet, while that is precisely where an automated pentest proves its worth.
What it means for your teams
The direction is interesting even if the tool still has to prove itself. For a three-person security team covering forty applications, a continuous pentest even an imperfect one beats a perfect annual test sitting in a PDF.
Points to check when evaluating this kind of service:
- Require real proof of exploitation in reports, not signature matches.
- Contain the credentials you hand over: dedicated accounts, limited scope, revocation after the test.
- Ask for false positive rates, or measure them on a pilot scope before rolling out.
The service joins the wave of security agents launched by major cloud providers, such as Cloudflare's security audit skill, and raises the same governance question as AI agent identity covered in RSA Agent ID: who supervises an agent that holds access rights on your systems?