AI writes your code. Can your security keep up? 

Security in AI-native, verified development: why the hybrid model is the way forward. 

AI now generates much of our code, but fixing it still runs at human speed. Here’s how a hybrid, verified model closes the gap.

The new reality: code arrives faster than security 

Roughly half of all organizations now generate at least part of their code with language models, and the volume they produce equals twice what a traditional team would develop. It is a genuine transformation of software, but it comes with an uncomfortable fact: AI-generated code contains almost three times as many cross-site scripting vulnerabilities as code written by people. It does not look like intentional behavior; it seems to be a consequence of how the models are configured and prompted. What is clear is that this code cannot be assumed to be secure by default. 

The problem is not just the speed of generation, but the speed of correction. Remediation is still a largely human, slow, and expensive process, and recent incident reports confirm it: more vulnerabilities are being discovered than can be fixed before applications are attacked. 

What about using AI to analyze and remediate? 

That is the obvious answer, and it works — up to a point. But those who have spent years applying AI models to code analysis have accumulated experience with three obstacles that cannot be ignored: 

  • Determinism. AI is probabilistic; security needs certainty. In tests run against an open, deliberately vulnerable reference application, a single model returned detection rates of 16%, 19%, and 35% on consecutive runs. The uncomfortable question asks itself: how many times do you have to ask before you trust the result? Thirty? And even then, there is no guarantee the list is complete. 
  • Costs. Every full analysis of an application consumes tokens, and in the same study each run cost about $29. After three analyses, only a third of the vulnerabilities present had been found. Multiplied by the number of applications and the sprint cadence of continuous development, the model does not scale economically. 
  • Governance. Audit, traceability, compliance: who reviews each finding, who closes it, what was done and why. AI produces answers, but not a governed history with an audit trail and the ability to reconstruct decisions months later — something indispensable in regulated environments or forensic investigations. 

Even the model vendors themselves acknowledge this in their documentation: coding assistants do not replace a security analysis platform, precisely because of this determinism problem. 

The hybrid model: native and verified AI 

The approach presented in the Spanish OpenText™ Fortify™ webinar combines both worlds instead of choosing one. On one side is traditional static analysis, which is deterministic. In Fortify’s case, it’s a base of 45,000 rules covering more than 1,600 known vulnerability families, the same result on every run, a fixed cost with no limit on the number of applications, and a governance platform with full history and audit trail. Its limit is well known: it does not capture problems that depend on business context or architecture, and on its own it does not produce the fix. 

On the other side is AI, capable of going one step further in detection and of proposing concrete remediations. The trick lies in the flow architecture: the deterministic analysis filters first, and AI only works on what remains. If the static engine detects 80% of the problems, you only pay tokens for the remaining 20%. The broader the coverage of the traditional engine, the smaller the AI bill and the greater the confidence in the whole. 

Three building blocks already in production 

OpenText Fortify Agentic Analyzer  

One more command inside the pipeline that runs the “agentic” analysis in three phases. Planning discards what is not worth sending to the AI, saving tokens from the very first minute. Analysis is performed by whichever model you choose. And validation cleans up the result: it removes false positives, drops duplicates already caught by the traditional rules, and packages everything in SARIF format to upload to the Fortify console, where it is governed alongside the rest of the findings. It is language-agnostic — supporting more than 46 — and compatible with over 1,600 models, local or cloud-based, open or commercial. It also lets you set a maximum cost per scan, a safeguard that matters when AI agents occasionally loop and send the bill through the roof. 

OpenText Fortify Aviator 

OpenText’s AI for audit and remediation, licensed per application rather than per token, with the capacity to resolve up to 32,000 vulnerabilities per application, from both static and dynamic analysis. It takes the Fortify findings, analyzes only the affected code fragment, and returns each vulnerability classified as exploitable, false positive, or suspicious — always with a reasoned explanation. For true positives it delivers the exact code to replace, with the specific lines and their location, so it can be applied without breaking the build. In the demo, Aviator processed more than 500 findings and left only about fifteen for prioritized human review. All of this integrates into the developer’s IDE, can be applied via pull requests inside the CI/CD flow, and even correlates static with dynamic findings: whatever shows up in both analyses deserves immediate attention. 

OpenText Fortify skills 

Capabilities you teach to any AI assistant — a coding agent, a conversational assistant — so it can work with the platform: generating compliance reports, creating custom Fortify static rules, or running an exploitability analysis on open-source dependencies. The last of these answers a classic question that is expensive to answer by hand: my application uses a library with a known CVE — but is it actually exploitable in my code? The agent generates a VEX report that distinguishes reachable components from unreachable ones, and can even apply patches only to the affected ones. In the example shown, only two of five vulnerable dependencies required prioritized fixing; the rest could be scheduled calmly. 

What the demo made clear 

Beyond the numbers, the hands-on part illustrated the value of context well. A static analysis may point to a line that, at first glance, is a harmless getter. The agent, by contrast, traces the full flow and discovers that the data is a credit card number that gets serialized and exposed through an API — that is where the finding becomes critical. This kind of reasoning about business flows is what AI brings to the hybrid model, and no fixed list of rules will ever fully cover it. 

The human role was equally clear. Every decision the AI makes — marking something as a false positive, confirming a remediation — is recorded in the audit trail alongside human decisions, and is reversible. If an auditor disagrees with the AI, they change the status and the history reflects everyone’s reasoning. And in automated remediation, the recommended flow ends in a pull request with a visible diff for review, although the platform can also apply changes directly. 

Checklist: where to start 

  • Inventory which code is already generated with AI and audit it with deterministic SAST. 
  • Do not replace traditional analysis with AI — complement it: rules first, AI second. 
  • Connect agents to CI/CD pipelines with per-scan cost limits and human review. 
  • Demand traceability: every AI decision must be recorded, explained, and reversible. 
  • Prioritize by real exploitability, not by the number of CVEs in your dependencies. 
  • Bring remediation into the IDE: developers fix where they write. 
  • Use static-dynamic correlation: a finding confirmed by both goes first. 

Conclusion 

AI already writes most of the code, and that trend will not reverse. The difference between an organization that survives this shift and one that suffers its consequences lies in how it reviews and fixes what AI produces. The hybrid model — deterministic analysis as the backbone, AI as the context and remediation layer, and governance from end to end — is not a roadmap: as presented, the pieces are already in production, and in the case of Agentic Analyzer and the Skills, included in the existing license. 

See the full webinar recording, including all the demos, to learn more. Please note: this recording is in Spanish.

Josef Hajda

Josef Hajda is a Field Marketing professional focused on driving regional engagement across the DACH market. He began his career in digital marketing, building a strong foundation in campaign execution and audience engagement. Since early 2025, Josef has been leading field marketing initiatives at OpenText, developing and executing regional programs and events that connect internal teams, partners, and customers to drive meaningful business outcomes.