Apex AI Pentesting Tool Finds App Vulnerabilities in Black-Box Mode

Key Takeaways

• PensarAI has introduced Apex, an AI-powered penetration testing agent designed for black-box vulnerability discovery in live applications.
• Apex operates without access to source code, hints, or predefined attack paths, aiming to keep pace with rapid AI-driven software development.
• To benchmark its capabilities, PensarAI developed Argus, an open-source suite of 60 Dockerized vulnerable web applications.
• During testing, Apex identified 271 unique vulnerabilities across the Argus challenges, achieving a 35% pass rate and outperforming other AI pentesting tools.
• Both the Apex agent and the Argus benchmark are now available as open-source projects on GitHub.

PensarAI Unveils Apex: An AI Agent for Black-Box Application Pentesting

In a significant development for application security, PensarAI has launched Apex, an autonomous, AI-driven penetration testing agent engineered to conduct black-box assessments against live applications. This innovative tool distinguishes itself by operating entirely without access to source code, architectural hints, or pre-configured attack vectors, enabling it to swiftly identify, chain, and validate real-world vulnerabilities at the speed of contemporary software development.

The impetus behind Apex stems from a recognized deficit in current software security practices. The rapid proliferation of AI coding agents, exemplified by Stripe’s agents merging 1,300 pull requests weekly and some engineering teams spending over $1,000 daily on AI tokens without human code review, has created an unprecedented velocity in code generation and deployment. Traditional security scanners and human-led penetration tests are often unable to keep pace with this accelerated development cycle.

Apex is positioned as an adversarial verification layer, designed to emulate a real attacker’s approach to an application, thereby catching exploitable weaknesses before they can be leveraged in a breach.

Flexible Deployment Across Development and Production

Apex offers versatile deployment options across different stages of the software lifecycle. In continuous integration (CI) pipelines, it can validate every deployment against a sandboxed replica of the application, systematically mapping the attack surface and attempting exploitation prior to code merges.

For production environments, Apex continuously uncovers exploitable weaknesses in real time. It also facilitates on-demand testing for any target, moving beyond the traditional quarterly PDF-based engagement to provide a continuous feedback loop aligned with the pace of modern threat landscapes.

Introducing Argus: A New Benchmark for Offensive Security Agents

To rigorously validate Apex’s capabilities, PensarAI developed Argus, an open-source benchmark comprising 60 self-contained, Dockerized vulnerable web applications. Argus was specifically engineered to evaluate offensive security agents, addressing perceived shortcomings in existing benchmarks.

PensarAI noted that widely used suites, such as XBOW’s 104-challenge set, are predominantly PHP-based (70%), focus on single-vulnerability targets, and lack coverage for modern technologies and attack scenarios like GraphQL, JWT algorithm confusion, race conditions, prototype pollution chains, WAF bypass, and multi-tenant isolation scenarios.

Argus, in contrast, encompasses frameworks prevalent in production environments, including Node.js/Express (40%), Python/Flask/Django (20%), multi-service architectures (25%), Go, Java/Spring Boot, and PHP. It introduces novel categories not covered by other benchmarks, such as WAF and IDS evasion, multi-step exploit chains requiring up to seven chained vulnerabilities, multi-tenant isolation failures, race conditions, business logic flaws, modern authentication bypasses (JWT, OAuth, SAML, MFA), and cloud/Kubernetes infrastructure attacks. The challenges are carefully calibrated for difficulty, with 2 easy, 27 medium, and 31 hard scenarios.

Benchmark Results: Apex Discovers 271 Vulnerabilities

During its evaluation, Apex was directed at all 60 Argus challenges in full black-box mode. For this initial run, it utilized Claude Haiku 4.5, selected as the smallest and most cost-effective model to highlight architectural efficiencies over raw model power.

Apex achieved a 35% pass rate, surpassing PentestGPT (30%) and Raptor (27%). When tested against the top 10 hardest challenges using Claude Opus 4.6, Apex significantly widened its lead, solving 80% compared to PentestGPT’s 70% and Raptor’s 60%.

Across the entire benchmark run, Apex successfully identified 271 distinct vulnerabilities, encompassing a wide range of types including SQL injection, SSRF, NoSQL injection, prototype pollution, SSTI, XXE, race conditions, IDOR, authentication bypasses, CORS misconfigurations, command injection, and path traversal. The average cost per challenge was approximately $8, with the total 60-challenge run on Haiku costing under $500.

Notable successes included uncovering a 7-step race-condition double-spend in a fintech transfer endpoint, exploiting a multi-tenant SSRF chain that pivoted through a shared cache to extract API keys from neighboring tenants, and achieving SpEL injection leading to Remote Code Execution (RCE) in a Java Spring Boot application—all within 15 minutes.

Analysis of Apex’s failure modes provided valuable insights. The primary challenge identified was last-mile execution, specifically the final step of credential extraction after a successful SSRF chain. Additionally, the agent was misled twice by decoy flags, and complex multi-step chains, such as CI/CD pipeline poisoning and Kubernetes compromise, exceeded the allocated 30-minute budget.

Both the Apex agent and the Argus benchmark are now publicly available as open-source projects on GitHub.

Disclaimer: HackersRadar reports on cybersecurity threats and incidents for informational and awareness purposes only. We do not engage in hacking activities, data exfiltration, or the hosting or distribution of stolen or leaked information. All content is based on publicly available sources.