The Taxonomy of Artificial Intelligence Security: Insights from the San Francisco [un]prompted Conference

The intersection of artificial intelligence and cybersecurity reached a critical inflection point recently in San Francisco, where industry leaders, academic researchers, and policy experts converged for the three-day [un]prompted AI security conference. Among the prominent voices at the dense, multi-track technical gathering was renowned security technologist Bruce Schneier, who delivered a keynote address focused on the principles of integrous artificial intelligence systems. The event underscored a growing consensus within the global technology sector that safeguarding artificial intelligence cannot be treated as a monolithic challenge. Instead, security experts are increasingly categorizing the discipline into a structured taxonomy: the security of artificial intelligence, security executed with artificial intelligence, and security required from artificial intelligence. This tripartite framework is rapidly becoming the foundational standard for how laboratories, enterprises, and defense organizations approach the next generation of digital defense systems.

Chronology and Event Background

The [un]prompted conference, held in downtown San Francisco over a rigorous three-day schedule, emerged in response to the escalating complexity of deploying machine learning models in high-stakes enterprise and governmental environments. As generative models, large language models (LLMs), and autonomous agent networks transition from experimental sandboxes into critical infrastructure, the threat landscape has expanded exponentially.

Leading up to the event, the cybersecurity community observed a dramatic surge in novel attack vectors targeting machine learning pipelines, including sophisticated prompt injection techniques, data poisoning, adversarial perturbation of training inputs, and model extraction attacks. Recognizing the fragmented nature of current research, the organizers of [un]prompted structured the summit to systematically address the full spectrum of vulnerabilities and defensive capabilities. The event featured concurrent technical workshops, peer-reviewed paper presentations, and high-level panel discussions that mapped out the trajectory of machine learning governance from late 2026 into the foreseeable future.

The Three Pillars of Artificial Intelligence Security

To effectively allocate research capital and engineer robust defenses, technology leaders at the conference emphasized the necessity of dividing the discipline into three distinct operational domains.

1. Security of Artificial Intelligence

The first pillar focuses inward, addressing the inherent vulnerabilities within machine learning models, training data pipelines, and underlying neural network architectures. This domain encompasses traditional software vulnerabilities compounded by the probabilistic nature of machine learning. Threat actors targeting the security of AI typically attempt to manipulate the model’s behavior during training or inference. Common challenges in this category include data poisoning—where malicious actors inject corrupted records into training sets to create backdoors—and adversarial examples, which subtly alter input data to cause catastrophic classification errors. Furthermore, privacy attacks such as membership inference and model inversion allow attackers to reconstruct sensitive training data from a model’s outputs, posing severe regulatory and compliance risks under frameworks such as the European Union Artificial Intelligence Act and various global privacy mandates.

2. Security with Artificial Intelligence

The second pillar shifts the paradigm outward, examining how machine learning systems can be leveraged as defensive multipliers to protect traditional software, networks, and cloud infrastructures. In this capacity, artificial intelligence acts as an automated security analyst. Modern Security Information and Event Management (SIEM) platforms, extended detection and response (XDR) tools, and automated orchestration platforms increasingly rely on neural networks to parse millions of telemetry events per second. These models excel at anomaly detection, identifying zero-day threat patterns, and accelerating incident response times far beyond human operational capacity. However, panels at the conference cautioned that relying on AI for defense introduces its own systemic risks, notably the susceptibility of defensive models to evasion techniques and the operational overhead of managing high rates of false positives.

3. Security from Artificial Intelligence

The final pillar addresses the macro-level systemic risks posed by the widespread deployment of advanced artificial intelligence capabilities across society. This domain extends far beyond technical cybersecurity into geopolitical stability, economic security, and public safety. Security from AI involves mitigating the risks of automated, highly scalable cyberattacks orchestrated by malicious actors using autonomous agents. It encompasses the proliferation of hyper-realistic deepfakes used for sophisticated social engineering and corporate fraud, as well as the potential misuse of dual-use foundational models for biological, chemical, or cyber warfare engineering. Addressing this pillar requires unprecedented international cooperation, robust regulatory oversight, and the establishment of verifiable safety alignment protocols before frontier models are released to the public.

Supporting Data and Industry Metrics

Recent empirical data presented at industry forums preceding the San Francisco summit highlight the urgent need for a structured approach to machine learning defense. According to enterprise risk assessments compiled throughout the technology sector, over 65 percent of organizations deploying generative AI tools into production environments have experienced at least one security incident related to prompt injection or unintended data leakage within the past twelve months.

Furthermore, capital expenditure dedicated specifically to artificial intelligence security tooling has grown by an estimated 140 percent year-over-year, reflecting corporate anxiety over regulatory penalties and intellectual property theft. Despite this influx of funding, independent security audits reveal that more than half of commercial large language model deployments lack basic input sanitization and output validation layers, leaving them vulnerable to indirect prompt injection attacks where malicious instructions are embedded within retrieved documents or external web pages.

Official Responses and Industry Reactions

Leading software enterprises, cloud service providers, and academic institutions have responded to the growing threat matrix by establishing specialized red-teaming units dedicated exclusively to machine learning infrastructure. Representatives from major technology firms participating in the San Francisco conference stressed that traditional cybersecurity frameworks—designed for deterministic, rules-based software architectures—are fundamentally inadequate for probabilistic systems.

"We cannot secure a neural network using the same static rulebooks we used for legacy relational databases and web servers," noted a senior security architect during a plenary session. "The industry must transition toward continuous validation, probabilistic auditing, and runtime guardrails that can dynamically evaluate model outputs in real time without degrading user experience or computational efficiency."

Regulatory bodies have also signaled approval of this segmented security taxonomy. Policy analysts attending the event noted that international standards organizations are actively incorporating these three distinct categories—defending the model itself, using the model for defense, and mitigating societal risks—into upcoming compliance frameworks. This structured approach provides a clear lexicon for legislators drafting compliance mandates, ensuring that technical requirements for model developers differ appropriately from the operational requirements of enterprise end-users.

Broader Impact and Strategic Implications

The formal articulation of the artificial intelligence security taxonomy marks a maturing of the cybersecurity profession. As artificial intelligence becomes deeply embedded in critical national infrastructure, financial networks, and healthcare systems, the margin for error narrows significantly.

The insights emerging from gatherings like the [un]prompted conference indicate that the future of digital defense will depend heavily on interdisciplinary collaboration. Cryptographers, machine learning researchers, sociologists, and policy makers must work in tandem to address vulnerabilities across all three fronts simultaneously. Neglecting the security of AI leaves individual enterprises vulnerable to espionage and data theft. Ignoring the security applications of AI misses an opportunity to automate and scale defensive postures against increasingly sophisticated cybercriminals. Failing to account for the security risks stemming from AI threatens global stability on an unprecedented scale.

As research continues to evolve and methodologies are refined, the adoption of a unified security framework will serve as the primary bulwark against emerging threats. The dialogue established in San Francisco provides a vital roadmap for the cybersecurity community as it navigates the complex, high-stakes transition into an artificial-intelligence-driven operational landscape.