State of (Absolute) AppSec

State of (Absolute) AppSec

Speakers: Seth Law, Principal Consultant, Redpoint Security; Ken Johnson, CTO & Co-founder, Dryrun Security; Kevin McDermott, Head of Security, Superhuman; Astha Singhal, Director of Security, Netflix; Clint Gibler, Head of Security Research, SERB

Conference: BSides SF

YouTube: https://www.youtube.com/watch?v=1H3zrcaKnH8

Overview

The "State of (Absolute) AppSec" panel at BSides SF delved into the seismic shifts occurring within application security due to the rapid advancements and integration of Artificial Intelligence (AI) into the software development lifecycle. Moderated by Ken Johnson and Seth Law, the discussion featured a distinguished panel comprising Kevin McDermott, Astha Singhal, and Clint Gibler, who offered diverse perspectives on the challenges and opportunities presented by this new technological era. The central theme explored how AI is not merely an additive tool but a fundamental disruptor, necessitating a radical rethinking of AppSec strategies, tooling, and professional competencies.

This talk is crucial because AI's impact extends beyond automating existing tasks; it's fundamentally altering how code is written, how applications behave, and where vulnerabilities emerge. The panelists highlighted the industry's tendency to underestimate the speed of this transformation and the unique security implications of non-deterministic software. The conversation provided a sobering yet optimistic look at the future, emphasizing that AppSec professionals must evolve from traditional compliance-focused roles to become "AI builders" capable of embedding security directly into the fabric of AI-native development.

The discussion underscored that while the core goal of reducing risk remains constant, the methods, processes, and even the types of vulnerabilities are undergoing profound change. From the automation of mundane tasks like DAST triage to the emergence of new architectural risks and the potential for AI to dramatically alter the OWASP Top 10, the panel offered a comprehensive overview of the current and future state of AppSec. It served as a call to action for the security community to embrace innovation, rigorous evaluation, and a proactive, engineering-centric approach to securing the next generation of software.

Background

▶ Watch: Panelist introductions for 'State of Absolute AppSec' (0:00)

For decades, application security practices have largely evolved around a relatively deterministic software development model. Whether adhering to waterfall or Agile methodologies, the premise has been that code behaves predictably, and security tools like Static Application Security Testing (SAST) and Dynamic Application Security Testing (DAST) are designed to identify known patterns and vulnerabilities within this predictable framework. The "shift left" movement and the rise of DevSecOps aimed to integrate security earlier into the traditional Software Development Lifecycle (SDLC), but these efforts often still conceptualized security as a series of checks and gates within a largely human-driven, linear process.

The advent of powerful AI, particularly large language models (LLMs) and generative AI, shatters many of these long-held assumptions. As Astha Singhal articulated, the industry is facing a "fundamental shift in how software will be built and written," demanding a re-evaluation of all traditional AppSec processes and controls. Kevin McDermott highlighted the profound impact of non-deterministic software, where AI components can produce unexpected outputs, challenging the very notion of "controlled, well-understood workflows" that traditional AppSec is built upon.

Furthermore, Clint Gibler pointed out that AI will dramatically increase the "amount of code that's going to be written and how quickly," alongside an accelerated "speed and ease of exploitation of new vulnerabilities." This rapid development and deployment cycle, coupled with the inherent unpredictability of AI, creates a significant disconnect with legacy AppSec tooling and audit requirements. The panel implicitly recognized that the existing "bolted-on" security approach, often driven by compliance rather than engineering innovation, is unsustainable in an AI-native world. The stage is set for a new paradigm, potentially an "agentic SDLC" or "ADLC," where AI agents are integral to both code generation and security enforcement, demanding a complete overhaul of how security professionals think and operate.

Key Findings

▶ Watch: Astha's take: Underestimating AI's fundamental shift in software (3:14)

The panel discussion yielded several critical insights into the evolving landscape of application security in the age of AI:

• Fundamental Shift in Software Development: Astha Singhal posited that AI represents a "fundamental shift in how software will be built and written," necessitating a complete re-evaluation of AppSec processes, context, and controls for an AI-native world. This includes moving towards an "agentic SDLC" or "ADLC" where AI is deeply embedded in every stage.
• Underestimated Impact of Non-Deterministic Software: Kevin McDermott emphasized that the industry is underestimating the impact of non-deterministic software. Traditional AppSec relies on predictable workflows, but AI introduces outputs that are not always controlled or understood, requiring new defensive strategies.
• Accelerated Code Generation and Exploitation: Clint Gibler highlighted that AI will drastically increase the speed of code generation and the ease of exploiting new vulnerabilities. Conversely, he noted that AI also empowers builders to create more secure-by-default libraries and "paved roads."
• Economic Disruption and Automation of Traditional AppSec: Seth Law observed a shift in consulting contracts and the need to justify manual AppSec services. He noted that tasks like secure code review, traditionally manual, can now be performed with "a single prompt" to AI agents, indicating a significant economic and operational disruption to traditional AppSec roles. Kevin McDermott's team, for instance, has automated "almost no human triage of DAST issues anymore" using AI.
• AppSec is Evolving, Not Dying: While traditional AppSec models may be "dead" in their current form, Astha Singhal passionately argued that "AppSec is not dead." Instead, she views it as the "most exciting time in application security ever," akin to being present at the invention of new vulnerability classes and solutions. The focus shifts to applying core competencies to new problem spaces.
• Widening Gap Between AppSec Teams: Kevin McDermott predicted a growing divide between security teams operating as engineering functions (adapting to AI) and those operating as compliance functions (sticking to traditional models). The former will be significantly more effective at reducing risk.
• Potential for Meaningful OWASP Top 10 Change: Clint Gibler suggested that AI could "meaningfully change the OWASP Top 10" for the first time in decades. By automating the mitigation of persistent issues like SQL injection and cross-site scripting through better coding models and secure-by-default libraries, future iterations could address entirely new categories of risk.
• Bug Bounty Programs as Entry Points: Clint also made a spicy prediction that bug bounty programs might "stop being a good entry platform for security people" within 6-12 months, as top bounty hunters or AI-powered products scoop up all low-hanging bugs, leaving junior researchers with "zero non-duplicate bugs."
• Criticality of Evaluations and Benchmarking: Ken Johnson, Astha Singhal, and Clint Gibler collectively stressed that evals and benchmarking are paramount for any AI-driven security solution. Without rigorous measurement and continuous refinement, AI tools risk becoming "this worked great at first" pitfalls. The need for a "data sciency mindset" in AppSec was emphasized.
• Talent Growth and Core Competencies: The panel agreed that future AppSec professionals will need strong "systems thinking and architecture" skills, curiosity, and the ability to "build" AI solutions, rather than just run scanners. The path into AppSec will still often start with development, but modern appdev fundamentals now include managing AI agents.

Technical Deep Dive

▶ Watch: Kevin on non-deterministic software impact on AppSec (4:00)

The panel extensively explored the technical ramifications of AI integration, highlighting both the capabilities of new tools and the emerging classes of vulnerabilities. The discussion moved beyond theoretical concepts to concrete examples of how AI is being leveraged and the technical challenges it introduces.

A significant theme was AI-driven automation within AppSec. Kevin McDermott provided a compelling, real-world example from Superhuman: his team has automated "almost no human triage of DAST issues anymore" by utilizing a Claude Code skill. This refers to an AI agent (likely based on Anthropic's Claude model) equipped with a specialized "skill" or function specifically designed to process DAST scan outputs, identify true positives, categorize them, and automatically generate tickets for remediation. This illustrates the power of agentic technologies to offload repetitive, time-consuming tasks previously requiring human expertise. Seth Law echoed this sentiment, noting that manual secure code review, which he and Ken Johnson once taught, can now be largely automated through "a single prompt" to AI agents, underscoring the shift towards programmatic security analysis.

Clint Gibler elaborated on how LLMs can integrate security much earlier in the SDLC/ADLC, specifically during the design phase. He described a scenario where an LLM could analyze design documents (e.g., Google Docs) written by developers or product managers. If the LLM identifies high-risk elements, such as interactions with sensitive databases or user login flows, it could automatically recommend specific authentication libraries or mitigating controls. These recommendations could then be "auto put into the requirements" for the project. Further, Clint suggested that after the code is built, a separate cloud code or coding agent could perform post-build validation, comparing the implemented code against the earlier security requirements to ensure adherence—effectively answering the question, "did you actually do the thing?" This vision represents a deeply embedded, continuous security feedback loop driven by AI.

However, the panel also cautioned against the blind adoption of AI. Seth Law highlighted a critical issue: while AI models can solve some security issues, they can also introduce new variants of old vulnerabilities. He observed a "resurgence of SQL injection" in AI-generated code, citing personal experience where upgrading from Claude 4.5 to 4.6 Opus led to PRs containing "literally SQL injection, basic SQL injection," because the model "completely ignored" previously installed security skills. This points to the need for continuous vigilance, the inherent fallibility of current models, and the importance of specific security guardrails. Astha Singhal further contextualized this, explaining that understanding prompt injection—where untrusted data is treated as code—requires fundamental systems thinking, akin to understanding traditional injection flaws.

The discussion also touched upon the broader ecosystem of security tools. Clint Gibler proposed an approach where AI models generate initial code, followed by hooks that trigger "autoscans that code with some carefully crafted prompts" using different AI agents. A post-processing step with "separate context that's not biased by what it just wrote" would then flag issues. This layered approach mirrors current human-driven code review processes, where multiple tools and human reviewers provide distinct perspectives to catch errors. This iterative refinement, combined with rigorous evals and benchmarking, is crucial for ensuring the reliability and effectiveness of AI-powered security solutions, preventing the "this worked great at first" scenario that Ken Johnson warned against.

Demo / Proof of Concept

▶ Watch: Clint on rapid code generation and easy exploitation (5:00)

The "State of (Absolute) AppSec" panel, by its nature as a discussion, did not feature a live, step-by-step demonstration or a detailed, visual proof of concept. The format was primarily an expert panel sharing insights, predictions, and experiences.

However, Kevin McDermott provided a compelling real-world example of an AI-powered solution in production at Superhuman. He explicitly stated that his team has achieved significant automation in their DAST triage process, noting, "We do almost no human triage of DAST issues anymore. It is all through a Claude Code skill that kicks off well an agent... with a skill to do all of our DAST triage for us." This direct application serves as a powerful testament to the practical capabilities of AI in AppSec, illustrating how agentic technologies can fully automate tasks that traditionally required significant human effort. While not a live demo, this concrete use case provided tangible evidence for the panel's overarching theme of AI transforming AppSec operations.

Defensive Implications

▶ Watch: Seth on AI's economic disruption in AppSec careers (6:00)

The insights shared by the panelists offer crucial guidance for AppSec professionals navigating the AI revolution. The overarching message is clear: adaptation, innovation, and a fundamental shift in mindset are non-negotiable for effective defense.

First and foremost, AppSec teams must re-evaluate and rebuild their programs from scratch, rather than attempting to retrofit legacy processes onto an AI-native world. As Kevin McDermott succinctly put it, "don't retrofit your current processes to AI data world," and "we need to get past that pretty quick and build actual systems that work as opposed to having the same paradigm but just AI it." This means embracing the concept of an agentic SDLC or ADLC, where security is intrinsically woven into every stage of AI-driven development.

Embracing automation is paramount. Repetitive, low-value tasks like DAST triage and basic secure code review are ripe for AI-driven automation. Kevin McDermott's example of using a Claude Code skill to automate DAST triage entirely demonstrates that such capabilities are not futuristic but achievable now. By offloading these tasks, AppSec engineers can free themselves to focus on higher-level strategic work, architectural risk, and securing the AI itself.

A proactive shift left is more critical than ever, with AI enabling deeper integration into the design phase. Defenders should leverage LLMs to perform design document analysis, identifying high-risk components early and auto-injecting security requirements, recommended authentication libraries, and mitigating controls. Subsequently, AI agents can perform post-build validation to ensure these controls are actually enforced in the generated code. This collaboration with platform engineering to build secure-by-default frameworks and components becomes a high-leverage point for embedding security.

The evolving vulnerability landscape demands new defensive strategies. While AI can help mitigate traditional flaws, the resurgence of issues like SQL injection in AI-generated code and the emergence of prompt injection require specific attention. Defenders must understand how these vulnerabilities manifest in AI contexts and develop appropriate controls, potentially including specialized AI agents for detection and remediation.

Perhaps the most significant defensive implication is the need for rigorous evaluation and benchmarking. As Ken Johnson, Astha Singhal, and Clint Gibler collectively stressed, any AI-powered security tool or process must be continuously measured and validated. AppSec teams need to adopt a "data sciency mindset," implementing robust evals and drift detection to ensure that AI solutions remain effective and do not introduce new blind spots or vulnerabilities as models evolve. Without this rigor, initial successes can quickly turn into unmanageable problems.

Finally, AppSec professionals themselves must upskill and re-skill. The future demands individuals with strong systems thinking and architecture skills, curiosity, and the ability to "build" AI solutions and integrate them into the development pipeline. The focus shifts from merely identifying vulnerabilities to actively constructing secure systems and managing the security of AI agents themselves. AppSec teams should transform into engineering functions, capable of rapid adaptation and continuous learning in a rapidly changing technological landscape.

Key Takeaways

• AI Reshapes AppSec Fundamentals: The integration of AI necessitates a radical shift from traditional, deterministic AppSec practices to an AI-native world and an agentic SDLC, where security controls are deeply embedded and continuously evolving.
• Automation of Mundane Tasks is Now Standard: AI is already automating tasks like DAST triage and basic secure code review, freeing AppSec professionals to focus on higher-value activities such as architectural risk assessment and building secure-by-default systems.
• Focus on Building and System Architecture: The future of AppSec lies in leveraging AI to build security directly into development pipelines, enforce secure defaults, and perform early design document analysis, requiring AppSec professionals with strong systems thinking and engineering capabilities.
• New and Resurgent Vulnerabilities Demand Vigilance: While AI can mitigate some traditional vulnerabilities, it also introduces new classes like prompt injection and can lead to the resurgence of old ones, such as SQL injection in AI-generated code, necessitating new defensive strategies and continuous adaptation.
• Rigorous Evaluation is Non-Negotiable: For any AI-driven security solution, continuous evals and benchmarking are critical to ensure effectiveness, detect drift, and prevent the introduction of new security blind spots, requiring a data-science approach to AppSec.
• AppSec Professionals Must Evolve: To remain relevant, AppSec teams must transition from compliance-focused roles to engineering-centric functions, emphasizing AI-building skills, curiosity, and a deep understanding of system architecture.

About the Speaker(s)

The panel brought together a diverse group of experts from leading security and technology companies, moderated by seasoned AppSec practitioners.

• Ken Johnson: As the CTO and Co-founder of Dryrun Security, Ken Johnson brings a wealth of experience in application security tooling and innovation. He co-moderated the panel, guiding the discussion on the future of AppSec.
• Seth Law: A Principal Consultant at Redpoint Security, Seth Law is also Ken Johnson's co-host for the "Absolute AppSec" podcast, which served as the inspiration for this live panel. He is deeply involved in secure code review and AppSec consulting, providing insights from the front lines of security implementation.
• Kevin McDermott: Head of Security at Superhuman, Kevin McDermott offered practical, real-world examples of AI integration into AppSec, notably his team's successful automation of DAST triage. His perspective highlights the operational challenges and successes within a fast-paced tech environment.
• Astha Singhal: Director of Security at Netflix, Astha Singhal provided a strategic, industry-leading perspective on the fundamental shifts AI is bringing to software development and security. Her background includes volunteering and running BSides, demonstrating a strong commitment to the security community.
• Clint Gibler: Head of Security Research at SERB, Clint Gibler brought a research-focused viewpoint, offering predictions on vulnerability trends and the broader impact of AI on the security landscape. His prior engagement with secure code review training, even years ago, underscores his deep roots in the field.