# AI That Hunts Logic Bugs Before Attackers Do

Taylor Luttrell-Williams, Sonar | AI Engineer NYC 2026 | 14:10

Source: https://www.youtube.com/watch?v=UW8Qkk7MG8A
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/ai-that-hunts-logic-bugs-before-attackers-do
Published: 2026-10-28
Tags: agents, multi-agent

## TL;DR
- SonarQube Hunter Agent looks for vulnerabilities caused by flawed intent, including broken access control, business logic gaps, and authentication failures that pattern-based SAST may miss.
- Hunter Agent maps a repository, runs three specialist agents, validates findings adversarially, and presents confirmed issues in SonarQube alongside SAST findings.
- In a Python application demo, Hunter Agent found 46 issues, including a user-controlled admin role and routes that did not require multi-factor authentication.

## Summary
Taylor Luttrell-Williams explains why pattern-based SAST does not catch every security problem. Code can be valid and still violate the application's intended rules, such as allowing a user to change their own role or reach an admin route without completing MFA. SonarQube Hunter Agent first maps the repository, then runs specialist agents for broken access control, business logic, and authentication or session issues. It validates leads adversarially, filters low-value findings, assigns severity, and adds confirmed issues to the existing SonarQube issue list. The demo scans a Python application in about an hour and reports 46 findings. Luttrell-Williams also shows scheduled scans, Sonar's partnership with the OpenAI Daybreak Cyber Partner Program, and a remediation agent that can automatically fix these issue classes. The talk's practical point is that Hunter Agent adds a second layer beside SAST and SCA for flaws that depend on how the application is meant to behave.

## Key ideas
### Valid code can still violate the application's intended security rules
[01:05](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=65s)
Taylor Luttrell-Williams distinguishes pattern-based SAST findings from intent problems. A program may be technically and functionally valid while contradicting the business logic behind it. Examples include logic-level vulnerabilities that arise from relationships between components rather than from a recognisable bad-code pattern. She says pentests may happen only twice a year or once a quarter, while agent-generated changes can ship much more often. Hunter Agent is intended to inspect those changes for flaws that ordinary SAST is not designed to find.

### Hunter Agent builds a repository-wide view before checking specific issue classes
[02:37](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=157s)
Hunter Agent first maps the repository globally. It then starts three specialist agents, with one for each class of issue it can find. The agents examine relationships between components and try to establish the application's security expectations. Hunter Agent follows this with adversarial validation to check whether a lead is a real issue rather than a false positive. Noise gates remove low-validity findings and issues that may be intentional. The resulting findings include severity, estimated impact, and the code location where developers can address them.

### Sonar is adding frontier-model reasoning through the OpenAI Daybreak program
[04:32](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=272s)
Luttrell-Williams announces that Sonar is joining the OpenAI Daybreak Cyber Partner Program. He says this gives products such as SonarQube Hunter Agent governed access to advanced cybersecurity models. The reasoning runs across the whole codebase, investigates every lead, and validates it before the result appears. He describes the resulting findings as precise, repeatable, and auditable. The program is presented as a way for security tooling to keep pace with attackers that also have access to increasingly capable models.

### Hunter Agent is intended to complement SAST and SCA in one issue workflow
[05:25](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=325s)
Sonar recommends using SAST and SCA together with Hunter Agent. SAST and SCA cover the checks they already perform, while Hunter Agent looks for the logic-level gaps left behind. The findings appear in the same SonarQube issue list, so users do not need a separate dashboard. Luttrell-Williams describes this as two independent layers of verification. The product can be used in SonarQube Cloud or in the self-hosted SonarQube Server variant.

### A full scan maps attack surfaces and consolidates specialist findings
[06:27](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=387s)
In the demo, Hunter Agent securely clones the target repository, analyzes its application architecture, maps possible attack surfaces, and builds a threat model. It then performs deep contextual scans for issues such as IDOR, broken access control, and business logic vulnerabilities. Findings from the specialist agents are aggregated and deduplicated into one issue list. Validated vulnerabilities are exported into SonarQube automatically. Luttrell-Williams says the demonstrated scan took about an hour, and a new scan can take longer depending on project size and complexity.

### The demo found 46 issues across access control, business logic, and authentication
[08:28](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=508s)
The Python application has 46 open findings across three categories. Broken access control includes IDOR, incorrect or missing authorization checks, privilege escalation, sensitive data exposure, and CSRF. Business logic findings include skipped workflow steps, repeatable actions being abused, and missing rate limits. Authentication and session findings include broken authentication, session fixation, sessions that do not expire, weak password recovery, missing MFA, and brute-force gaps.

### A user-controlled role field can bypass the admin check
[10:01](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=601s)
One high-severity finding shows that any signed-in user can pass the admin access check because the user can write their own role field. Hunter Agent points to the location in the code, explains why the behavior is a vulnerability, and suggests how to fix it. The issue can be assigned to a developer. Users can accept a valid issue without letting it affect quality gates, confirm it so it affects the quality gate, or mark it as a false positive.

### Scheduled scans and automatic remediation extend the workflow
[12:26](https://www.youtube.com/watch?v=UW8Qkk7MG8A&t=746s)
Hunter Agent can run on demand or on a schedule. The schedule can be daily, weekly, or monthly, and users can choose a branch instead of scanning the main branch. Luttrell-Williams also announces that the issues can be assigned to the SonarQube remediation agent. That agent can remediate these security issue classes automatically and can operate on a schedule.

## Notable quotes
- "These are issues of intent, issues related to logic-level vulnerabilities, wherein the code may very well be valid, but the intent itself is not." (01:30)
- "We then run adversarial validation to make sure that the issues are actually correct and not false positives." (02:52)
- "You can see here that there are 46 open issues on this project broken across the three issue categories that I touched on earlier." (08:24)
- "Any signed-in user being able to pass the admin access check because they can write their own role field." (10:38)

## Tools & references mentioned
- SonarQube
- SonarQube Hunter Agent
- SonarQube Cloud
- SonarQube Server
- OpenAI Daybreak Cyber Partner Program
- SAST
- SCA
- IDOR

## Who should watch
- You maintain applications where authorization, workflow rules, or MFA behavior cannot be checked reliably by pattern-based SAST alone.
- Your team already uses SonarQube and wants repository-wide checks for logic-level security issues in the same issue workflow.
- You need recurring scans or automated remediation for findings that can be assigned to developers and tracked through quality gates.

## Related talks

- [AI's Jurassic Park Period](https://aietalks.com/talks/ais-jurassic-park-period) (Aaron Stanley, dbt Labs, 21:42)
- [AI Hackers Are Faster Than Your Pen Test](https://aietalks.com/talks/ai-hackers-are-faster-than-your-pen-test) (Eli Cohen, Snyk, 19:29)
- [How We Hacked YC Spring 2025 Batch's AI Agents](https://aietalks.com/talks/how-we-hacked-yc-spring-2025-batchs-ai-agents) (Rene Brandel, Casco, 17:33)
- [Your agent is blindfolded](https://aietalks.com/talks/your-agent-is-blindfolded) (Johan Lajili, Poolside, 09:58)
- [Your AI Isn't the Problem. Your Infrastructure Is.](https://aietalks.com/talks/your-ai-isnt-the-problem-your-infrastructure-is) (Jiachen Jiang, Coder, 31:30)
