Trident, a security startup backed by Y Combinator, launched an autonomous penetration testing platform in August that scans cloud infrastructure and applications for vulnerabilities, then delivers findings as pull requests. The company chains exploits across Amazon Web Services, Microsoft Azure, Google Cloud Platform and Kubernetes into complete attack paths, a capability it says fills gaps left by traditional manual security audits.
Most organizations commission penetration tests once or twice annually, leaving infrastructure unexamined for months as engineers ship new code and adjust configurations. Trident's platform runs on a continuous schedule, surfacing attack sequences that cross cloud boundaries. A compromised CI identity in one provider might hold credentials granting access to another—precisely the kind of lateral movement that annual point-in-time audits tend to miss, according to the company's website.
The product builds what founders Junaid Mahmood and Dekai Li describe as "one graph across AWS, Azure, GCP and Kubernetes" that maps routes to sensitive data stores. Trident correlates IAM permissions, exposed secrets and scanner findings into what it calls "toxic-combination paths," then identifies choke-point fixes where a single remediation breaks multiple attack chains.
Every flagged vulnerability sits in a "Validating" state until the platform's agents confirm they can reproduce the exploit end-to-end. Confirmed findings ship with a replay request and a draft pull request containing a regression test. "Model output on its own is never treated as proof of a vulnerability," the company states in its FAQ, a nod to skepticism around AI-generated security alerts that never get validated in practice.
Trident's agents probe authentication boundaries, session handling, insecure object references and business logic flaws in web applications and APIs, aligning with OWASP WSTG. The focus skews toward issues that signature-based scanners overlook. The platform hooks into 19 integrations spanning cloud providers, identity systems like Okta and AWS IAM, data warehouses including Snowflake and BigQuery, and DevOps tools such as GitHub and Terraform.
It also exposes an MCP server and command-line interface so coding assistants—Claude, Cursor, GitHub Copilot—can pull findings and verify fixes directly.
The system runs scheduled attacks against customer infrastructure, which the terms of service flag as "the most important section, because Trident actively attacks the systems you point it at." Safeguards include allowlists, rate limits and non-destructive payloads. Many teams start in staging environments before enabling constrained production scans, according to the web application documentation.
Trident processes source code and credentials transiently, storing only the findings themselves. "We don't use your code, cloud or repository contents to train AI models," the terms state, effective since late August. The startup's testing agents run on foundation models from OpenAI and Anthropic. Sensiba LLP audited the company for SOC 2 compliance, with reports available upon request.
The founders operate Trident under Esprit Labs Inc., a Delaware entity. Y Combinator backed the startup in the S26 batch, 2026. The product went live with free access requiring no credit card; paid tiers add export and retention features.
According to industry observers, Trident enters a crowded field. Invicti announced an agentic penetration testing product in July combining autonomous reasoning with proof-based dynamic application security testing. Cobalt launched Autonomous Pentest targeting 24-hour turnaround times that same month. YesWeHack introduced its agentic tool in June, naming French enterprise customers including Dassault Systèmes. AWS shipped a multi-agent penetration testing capability in April as part of its Security Agent offering.
Horizon3.ai's NodeZero has been testing infrastructure autonomously since 2022, with documentation updated regularly through this year.
Academic researchers have flagged risks inherent in AI-powered penetration testers. A September paper on arXiv argued that generic guardrails prove insufficient for agentic offensive security tools and proposed a taxonomy of attacks targeting the agents themselves—a reminder that the weapons can be turned on their wielders.
Trident declined to disclose customer names or headcount as of late September. The company has published research on its blog in June and July 2026, including an analysis of Instagram account takeovers and a vulnerability disclosure concerning a cryptocurrency vault. The startup wrapped up Y Combinator's Demo Day in early September and currently offers free trial access through its site.
