DKDKCISSP
AI SecurityINCIDENT

Too many instructions can hold back AI security audits

GitHub's AI security audits face challenges due to excessive instructions in prompts.

DKCISSP News Desk17 Nov 2026, 12:30 pm2 min read
Too many instructions can hold back AI security audits
Image courtesy of Help Net Security. Original report

Peter Stöckli, a Security Researcher at GitHub, discussed the use of AI agents for auditing software vulnerabilities, highlighting that too many instructions can hinder their effectiveness.

In a recent interview, Peter Stöckli of GitHub explained how the company employs AI agents to audit software for vulnerabilities, focusing initially on authentication and authorization bugs. This approach stems from the understanding that many vulnerabilities arise from inconsistencies in access checks across different endpoints. Stöckli emphasized that while AI can effectively identify these issues, the quality of prompts used to guide the agents is crucial for accurate results.

The process begins with the AI agents analyzing various components within a software application. During threat modeling, the agents assess all entry points, including functions and web APIs, to capture the intended behavior of the software. A method that executes commands is benign if intended, but it becomes a vulnerability only if unintended data can exploit it. This nuanced understanding is key to how the AI identifies potential security flaws.

AI security audits involve sophisticated technology where machine learning models analyze code and configurations to identify vulnerabilities. These models, particularly large language models (LLMs), excel at recognizing patterns and inconsistencies, which are often indicative of security risks. However, the effectiveness of these models is heavily dependent on the prompts provided to them. If prompts are overly complicated or restrictive, they can stifle the AI's ability to generate creative solutions or accurately identify issues.

GitHub's AI auditing process includes a human review stage, where security researchers validate every report generated by the AI. This review process not only filters out false positives but also ensures that the reported vulnerabilities are significant enough to warrant attention. Each report comes with a proof of concept to aid maintainers in understanding the issue, fostering collaboration between the AI team and software maintainers to address vulnerabilities effectively.

The scale of this initiative is substantial, as GitHub aims to enhance the security of countless software projects hosted on its platform. By focusing on high-impact vulnerabilities and improving the reporting process, GitHub seeks to maintain a high fix rate for identified issues, which Stöckli noted is currently at 95%. This success rate is attributed to the combination of AI capabilities and human oversight.

While GitHub has made significant strides in automating vulnerability detection, the company continues to face challenges in refining its approach. Stöckli mentioned that the balance between providing sufficient guidance to the AI while allowing for creative problem-solving is delicate. The ongoing evaluation of prompts is essential to ensure that the AI remains effective without being bogged down by unnecessary instructions.

Looking ahead, GitHub plans to introduce more features to streamline the reporting and triage processes for maintainers. Enhancements such as structured reporting forms and API improvements are designed to automate some of the manual work involved in addressing vulnerabilities. However, the specifics of these upcoming features and their potential impact on the auditing process remain to be seen.

MORE IN AI SECURITY

More cybersecurity reporting

Your AI red team may be finding the same prompt injection thousands of times20 Nov 2026, 12:00 pmAI agent API security hasn't caught up with agents already running in production20 Nov 2026, 11:00 amPalo Alto Networks to acquire Applied Compute for $4.5 billion19 Nov 2026, 3:53 pmAI agents inherit employee access, and nobody's watching how they use it19 Nov 2026, 10:30 am