Pentest Superintelligence represents a new stage in the evolution of penetration testing, succeeding the Pentest AI-First and Pentest AI-Native methodologies we developed previously. This new model combines multiple high-capability frontier artificial intelligence models, orchestrated by a proprietary runtime, to run pentests with autonomy, technical depth and specialist validation.
The main differentiator lies in the architecture used to conduct the attacks. Instead of depending on the capability of a single artificial intelligence model, Pentest Superintelligence allows different models to work in a coordinated way during the same offensive operation, sharing context, exploiting vulnerabilities and developing attack chains. To make this possible, we developed a proprietary harness integrated into Yaga, our autonomous pentest agent.
The evolution from Pentest AI-First and AI-Native
In March 2026, we developed Pentest AI-First, a methodology created to place artificial intelligence in the first layer of the offensive operation. In that model, an AI agent performs reconnaissance, enumeration, contextual analysis and initial exploitation of vulnerabilities, while human specialists validate the findings and go deeper into the more complex attacks.
The proposal was born from the need to overcome the limits of traditional automation, demonstrating that running scanners and predefined scripts is not the same as performing a pentest. A penetration test requires offensive reasoning, adaptation to the environment, business logic analysis and the ability to chain different vulnerabilities together to prove real impact.
Later, in June 2026, we evolved that approach with Pentest AI-Native, significantly expanding the autonomy of artificial intelligence. In that model, Yaga began conducting the pentest end to end, from initial reconnaissance through vulnerability exploitation, impact validation and the organization of technical evidence, with our specialists following and validating the results.
The next challenge became expanding the offensive reasoning capability of the agent itself. After all, artificial intelligence models have different capabilities and can show distinct performance depending on the nature of the vulnerability, the complexity of the environment and the stage of an investigation.
It was from that need that we developed Pentest Superintelligence, a new architecture that makes it possible to draw on different frontier models during the same penetration test.
What Pentest Superintelligence is
Pentest Superintelligence is a new pentest model in which an autonomous artificial intelligence agent uses two or more high-capability frontier AI models, integrated and coordinated by a proprietary runtime, to conduct a complete offensive operation within an authorized scope.
This is the main difference from Pentest AI-Native. While AI-Native establishes the agent's autonomy to execute penetration tests, Superintelligence adds an architecture capable of orchestrating multiple intelligences during the execution of those tests.
For an operation to meet the technical criteria we established for Pentest Superintelligence, there must be an architecture with at least two high-capability frontier models, an autonomous agent and a proprietary runtime able to coordinate those models effectively during the same pentest.
This means that simply connecting different AI models to an application does not characterize this new model. The architecture must allow the models' capabilities to be used in a complementary way, preserving the context of the findings, making decisions during the operation and maintaining continuity across the exploitations.
The goal is for different intelligences to work as part of a single offensive operation, not as independent systems running isolated tasks.
How Yaga's architecture works
To implement Pentest Superintelligence, we developed a proprietary harness designed specifically for offensive cybersecurity operations. This harness works as Yaga's execution and orchestration layer, making it possible to coordinate different frontier models, including technologies from companies such as OpenAI, Anthropic and xAI.
During a pentest, Yaga can use different models according to the needs of each stage. One model may offer greater capability in interpreting code and analyzing business logic, while another may contribute to planning exploitations, investigating unexpected behavior or building more complex attack chains.
The harness is responsible for deciding which model should lead a given stage, providing the necessary context and incorporating the results obtained back into the ongoing operation. In this way, a finding produced by one model can be used by another to deepen the investigation, without the process having to start over.
This architecture allows Yaga to perform technical reconnaissance, formulate hypotheses, run real exploitations, investigate recently disclosed vulnerabilities and correlate findings to identify attack paths with greater impact.
Another fundamental element is execution control. The runtime operates with guardrails, limits on what it may do and control mechanisms compatible with the scope authorized by the client. Evidence is organized in a reproducible way, allowing our specialists to follow the findings, validate the vulnerabilities and go deeper into the attacks when necessary.
The result is a single pentest conducted by an autonomous agent, using multiple high-capability AI models and an offensive infrastructure developed specifically for that purpose.
The results of Pentest Superintelligence
We also developed YagaBench, our internal benchmark created to evaluate Yaga's offensive capability and to compare the results of its proprietary architecture with artificial intelligence models run individually.
In the evaluation published in October 2026, Yaga, using six orchestrated AI models, reached the following results:
- 98% resolution of the Black Box scenarios.
- 98.7% resolution of the Gray Box scenarios.
- 99% resolution of the White Box scenarios.
In those same Black Box scenarios, the best-performing individual model reached 40.5% resolution, while Yaga reached 98% using its orchestration architecture.
The benchmark uses a fixed set of offensive scenarios, with vulnerabilities effectively exploited and confirmed by our specialists. Yaga's runs were also executed with the same safety and control mechanisms used in client operations.
These results demonstrate the potential of combining frontier models with a specialized offensive execution architecture. More than the individual capability of an artificial intelligence, a pentest agent's performance depends on how it uses tools, preserves context, conducts investigations and turns findings into real exploitations.
The complete results and the evaluation methodology are available at YagaBench.
Offensive artificial intelligence with human validation
The evolution toward Pentest Superintelligence also expands the role of offensive cybersecurity specialists within the operation.
In our methodology, Yaga conducts the technical execution of the pentest, while our specialists follow the tests, validate every vulnerability identified and go deeper into the attack paths that require additional investigation.
This model makes it possible to concentrate human experience on activities of higher technical value, such as analyzing business-specific risks, interpreting complex impacts and investigating scenarios that call for specialist decisions.
It also makes it possible to increase the frequency of penetration tests, keeping pace with corporate environments that receive constant updates, new integrations, infrastructure changes and features built with artificial intelligence.
The priority remains the same: identify real vulnerabilities, technically prove their impact and deliver evidence that allows companies to fix their security risks.
A new model for the future of offensive cybersecurity
Pentest Superintelligence represents the third stage in the evolution of the pentest methodologies we have developed.
Pentest AI-First established artificial intelligence as the first layer of the offensive operation. Pentest AI-Native made it possible to run complete pentests through autonomous agents. Now Pentest Superintelligence expands that capability by integrating multiple frontier models within a proprietary execution and orchestration architecture.
This evolution also changes the way we evaluate the application of artificial intelligence to cybersecurity. The differentiator is no longer only the capability of an isolated model, and now involves the entire architecture that turns intelligence into real offensive results.
With Pentest Superintelligence, we have set out our vision for the next generation of penetration testing: autonomous agents, multiple frontier artificial intelligence models, proprietary runtimes and human specialists working together to expand the depth and execution capability of pentests.
The future of offensive cybersecurity demands more than using artificial intelligence. It demands building the technology capable of turning different intelligences into a single offensive operation.
Discover Pentest Superintelligence and see how this new architecture is being applied through Yaga and the HAS platform.