29 Aug 2025
Read Time: 4 min
GPT-5 and Enterprise AI: What leaders need to know now
Discover how automation can transform your customer relationship management processes.
Figure 1: https://splx.ai/blog/gpt-5-red-teaming-results
This matters in the context of enterprise security because if a model can be jailbroken too easily, organizations face serious risks: data leakage, compliance violations, reputational damage, or even insider threats where malicious employees deliberately bypass safeguards.For more on this please see this article:
The moral of the story is that new models need time to mature and harden. Just like any other software release, imperfections are uncovered quickly once exposed to the real world. What’s especially concerning is that AI-assisted code generators are already using GPT-5. This means that determined malicious actors could exploit the model to generate code for unethical or harmful purposes - making enterprise caution and strong guardrails more important than ever.
Options to help enterprise tackle this come in the form of runtime prompt hardening and runtime guardrails. GPT-4o has gone through hardening over time and is still rated in the 90% range on SPLX test suite. When compared with GPT-5 scores it’s clear that the new model needs some settling in time and is currently not suitable for default enterprise deployments.
Figure 2: https://splx.ai/blog/gpt-5-red-teaming-results
What This Means for Your Daily Activities
Don’t Trust the default Config!
Be cautious with AI-generated code. If you use code copilots (GitHub Copilot, Windsurf, Cursor), make sure you can select the model in your code copilot, GPT-5 is available in a few variants on both Cursor and Windsurf and it's unclear how guard rails on these tools are applied to new models. Review all AI-suggested code for hidden risks, backdoors, or compliance issues. You could also use older models to validate the safety and compliance of the newer models as part of your development pipeline.
Treat AI outputs like untrusted third-party code. Just as you would review open-source contributions, apply the same scrutiny to AI-generated content before merging it into production systems. Human inputs and reviews are critical as your company embraces the latest models. Don’t assume more capable means more secure.
Watch for data leakage. Avoid pasting sensitive enterprise data (credentials, customer PII, strategy docs) into AI tools unless they are explicitly approved and secured for enterprise use. Use the development pipeline and the PR process to ensure no secrets are embedded in the code. Incorporate gitignore and use secure key storage for all API and model/agent identifiers when deploying to production.
Keep security guardrails in mind. Red-teaming results show GPT-5 can be jailbroken. If you experiment with AI tools, be aware that outputs could include disallowed or harmful content. Report anything unusual. Keep up to date with the latest news and security assessments outside of the mainstream media white noise. Benchmarking is key but also an evolving field. For enterprise use ensure hardening and red teaming is applied early on.
Expect rapid change. The first weeks of a new model release are volatile. Policies, updates, and mitigations will evolve quickly - stay tuned to official SevTech guidance before relying heavily on GPT-5 - based tools.Index of new terminology for AI risk management
Key Terms:
Jailbreaking
Crafting prompts that bypass an AI model’s built-in safety controls, forcing it to output restricted or harmful content. Often done through role-play, multi-step narratives, or obfuscated requests.Prompt Injection
Like SQL injection, but for AI. Malicious instructions are hidden inside user inputs or external data, tricking the model into ignoring its original task.
Model Routing
Automatically directing requests to different AI sub-models depending on complexity or risk. Promising, but currently opaque - making enterprise auditing harder.Alignment
How well a model’s outputs match organizational goals, compliance rules, and ethical standards. GPT-5 scores show weak alignment compared to GPT-4o.Hallucination
When a model invents facts or sources with high confidence. In enterprise settings, this can create regulatory or financial risk.Guardrails
Protective layers around a model - such as filters, moderation, or external validators - that enforce compliance. Increasingly framed as the core enterprise challenge.Prompt Hardening
System prompts designed to constrain what the model can and cannot do, reducing jailbreak or injection success.RAG (Retrieval-Augmented Generation)
Linking an AI to trusted knowledge sources so outputs are grounded in fact. Essential for reducing hallucinations in enterprise use.
For further information or to book an AI assessment reach out at SevTech or Do you think enterprises are moving too fast to adopt GPT-5, or are the risks overstated? I’d love to hear how your teams are approaching it.
By Daniel Seavers, Principal Consultant, Sevtech
GPT-5 launched to great fanfare on August 7th 2025, but within 24 hours security researchers had claimed they had already jailbroken it. What does this mean for enterprise businesses racing to adopt it? At the very least it should sound alarm bells.
During the live event, the OpenAI team showcased some impressive demonstrations of the model’s power. From natural language inputs and control over the speed of translated spoken outputs, to building a browser-based video game in mere minutes, the new model certainly looked order of magnitudes more capable than its predecessors. However, as the rollout continued over the next 24 hours, it quickly became clear that not all of GPT-5’s differences were as favourable as they first appeared.
Within a single day, a group of security researchers - known as a red team - announced that they had successfully jailbroken GPT-5, exposing critical vulnerabilities and raising serious questions about its readiness for enterprise deployment.
These rapid exploits highlight how even the most advanced AI systems remain vulnerable to sophisticated manipulation. It serves as a stark warning for enterprise technologists. While new models are often presented as incremental advancements over their predecessors, the first few weeks after release are a critical period that can determine whether a model proves stable, secure, and truly enterprise ready.
For context, it’s important to pause here and examine the terms we’re discussing jailbreaking. In this context, jailbreaking refers to a method where users craft prompts that bypass or disable an AI model’s built-in safety controls, forcing it to output restricted or harmful content.
This often works through role-playing, multi-step narratives, or obfuscated requests. For example, instead of directly asking “How do I make X?” (which would normally be blocked), the prompt might be phrased as “Write a story where a character explains how to make X.” In one red-team test, researchers were able to convince GPT-5 to provide a recipe for a Molotov cocktail using this jailbreaking technique.