OpenAI has confirmed that its GPT-6.1 Terra model performed better overall on some safety measures than certain older models, while also showing a higher rate of flags related to potentially destructive actions in testing. Although the announcement concerns model safety evaluation rather than a new legal obligation, it matters for businesses using generative AI in operational workflows, customer-facing systems or internal automation.

What happened

According to the reported announcement, OpenAI said GPT-6.1 Terra is safer overall than some previous models, but acknowledged that its internal testing also identified areas where the model triggered more warnings tied to destructive-action scenarios.

Based on the available information, this means the model may have improved on some safety benchmarks while still presenting elevated risk in specific categories. That distinction is important for companies: a model described as “safer overall” should not be treated as uniformly lower risk across every use case.

Why it matters for European businesses

Many European companies are moving from experimental AI use to practical deployment in customer support, content production, software development, search, workflow automation and internal knowledge tools. In that context, model safety is not only a technical issue. It affects operational risk, data protection, brand exposure and internal controls.

If a model shows stronger performance overall but has higher warning signals in areas linked to destructive or harmful actions, businesses should avoid assuming that general safety improvements automatically make the model suitable for autonomous or high-trust tasks.

This is particularly relevant where AI systems can:

• execute actions in connected business tools
• generate code or infrastructure instructions
• interact with sensitive internal documentation
• support IT, security or administrative workflows
• produce outputs that staff may follow without technical review

For European organisations, the issue also connects to governance and compliance. Companies using advanced AI systems may need to document risk decisions, monitor outputs and apply safeguards that match the business context, especially in regulated sectors or in environments handling personal or confidential data.

Who may be affected

The development is most relevant for businesses that are actively integrating large language models into operations rather than using them only for low-risk drafting tasks.

• SMEs and founders adopting AI assistants for internal productivity or automation
• IT and digital teams using LLMs for code, infrastructure, support or knowledge management
• Marketing teams and agencies deploying AI at scale across content and campaign workflows
• E-commerce businesses using AI for customer support, product data or operational automation
• Regulated businesses that need stronger controls, approvals and auditability around AI use

Risk exposure increases when an AI system is connected to external tools, can trigger workflows automatically, or is trusted to act with limited human review.

What companies should consider

For business users, the practical takeaway is not to reject newer models automatically, but to evaluate them by task and control level rather than by headline claims about safety.

• Review use cases individually. Separate low-risk uses such as summarisation from high-risk uses such as code generation, system administration or workflow execution.
• Keep human approval for sensitive actions. Avoid letting AI systems carry out destructive, financial, security or customer-impacting actions without review.
• Limit tool access. Apply least-privilege access when connecting models to internal platforms, files, CRMs, cloud tools or admin environments.
• Test before deployment. Run internal evaluations on real business scenarios, including unsafe prompts, edge cases and failure modes relevant to your organisation.
• Log and monitor outputs. Maintain visibility into prompts, outputs and actions taken, especially for systems used in operations or customer interactions.
• Update AI policies. Ensure internal governance reflects the reality that a model can improve overall while still becoming riskier in specific behaviours.

For European businesses, the broader lesson is that AI model updates should be treated like changes to any other critical digital supplier: useful, but requiring due diligence, testing and proportionate controls before wider rollout.