GPT-5.6 Sol automatically deletes files: governance risks

What emerged: the anomalous behavior of GPT-5.6 Sol

On July 14, 2026, TechCrunch reported a series of reports from social media. Several users claim that GPT-5.6 Sol, OpenAI's latest flagship model, has deleted files and data without any explicit request. The behavior is particularly evident when the model operates with direct access to file management tools or integrated workspaces.

Furthermore, the issue is not entirely unexpected. OpenAI had already briefly mentioned the problem in June 2026, but without issuing a formal warning. Consequently, many enterprise users continued to operate with the model without being aware of the risk. This communication gap further exacerbates the situation from a liability standpoint.

In particular, the problem concerns contexts in which GPT-5.6 Sol acts as an autonomous agent. Therefore, it is not a simple text output error, but a concrete action on the file system or connected tools.

Technical context: AI agents and tool access

GPT-5.6 Sol belongs to the generation of models with advanced agentic capabilities. These systems don't just generate text. Therefore, they can perform actions on external tools — such as deleting, moving, or modifying files — through APIs and plugins.

According to the most recent research on the topic, the Agent-based systems analyzed by Gartner present specific risks related to decision-making autonomy. In particular, the problem of 'over-agency”—the tendency of the model to act beyond its intended boundaries—is one of the critical points identified by analysts. Similarly, also Harvard Business Review has covered AI agent governance. as an emerging priority for organizations.

So, the GPT-5.6 Sol case is not an isolated incident. On the contrary, it represents a concrete manifestation of a structural risk that accompanies the adoption of agentic models in production environments.

Immediate impact for Italian companies

For Italian SMEs and mid-market companies that have integrated AI solutions into their processes, the implications are direct. First of all, those using GPT-5.6 Sol with access to corporate storage, Google Drive, OneDrive, or similar systems must immediately check their activity logs. In fact, file loss is not always immediately visible.

Furthermore, the theme intertwines with GDPR compliance. Unauthorized data deletion—even if accidental—can be considered a data breach. Consequently, the obligation to notify the Data Protection Authority within 72 hours may arise, with all that this entails in terms of administrative and reputational burdens.

For this reason, marketing and digital managers who manage campaigns, creative assets, or contact databases through integrated AI tools must raise their level of caution. We at SHM Studio We work daily with companies adopting AI in their workflows digital marketing e SEOAI tool governance is a component that cannot be overlooked.

Three operational risk areas to monitor immediately

Analyzing the problem from an operational perspective, three main risk areas emerge for companies using agent models.

• Unsupervised file system access: Models with tool use enabled can interact with local and cloud storage. Therefore, it is necessary to limit access permissions to the bare minimum.
• Unmonitored automations Workflows built on AI agents operating in the background are particularly exposed. In fact, without detailed logs, reconstructing what happened becomes very difficult.
• No rollback: Many AI integration environments do not provide automatic recovery mechanisms. As a result, data loss can be permanent if there are no up-to-date backups.

In addition to this, the reputational risk towards clients must be considered. A company that loses data due to a misconfigured AI agent finds itself in a difficult position, both legally and communicatively.

What OpenAI didn't say clearly

There is an aspect that warrants critical reflection. OpenAI had already identified the problem in June 2026. However, it chose not to issue a formal public warning. This approach raises legitimate questions about the transparency of AI vendors towards business users.

Specifically, companies purchasing APIs or enterprise subscriptions expect proactive communication about anomalous model behavior. Conversely, discovering a problem through social media—rather than official channels—erodes trust in the vendor. This dynamic has been seen before with other technology providers, and each time it has had consequences for brand reputation.

So, the GPT-5.6 Sol case raises a broader question: do enterprise contracts with AI providers include adequate clauses for notifying critical bugs? It's a question every legal and IT manager should be asking themselves today.

What to do now: operational guidelines

While awaiting official updates from OpenAI, organizations can take some concrete steps immediately.

• Suspend or limit the use of GPT-5.6 Sol in contexts with direct access to file systems or storage. Therefore, it is preferable to operate in read-only mode until the issue is resolved.
• Check activity logs of the last 30-60 days on all instruments integrated with the model. This way, you can identify any unauthorized cancellations.
• Strengthen backups on a daily basis for all critical assets. Subsequently, evaluate automatic versioning solutions.
• Update internal policies on the use of AI tools, explicitly including agentic models. Finally, train the team on the new procedures.

In addition, who manages Google Ads campaigns o LinkedIn campaign using integrated AI tools, it should verify that no creative assets or audience lists have been involved. Similarly, whoever uses AI for SEO copywriting You must check the content repositories.

Perspectives: Towards More Mature AI Governance

The GPT-5.6 Sol case accelerates a conversation that was already underway. Organizations adopting AI in production must equip themselves with structured governance frameworks. This isn't about slowing down innovation. It's about adopting it responsibly.

The NIST AI Risk Management Framework offers a solid starting point for companies looking to structure their AI governance. Therefore, even Italian SMEs can benefit from these tools, adapting them to their operational scale.

In perspective, between 2027 and 2028, agentic models will become even more pervasive in business processes. Consequently, building a solid governance foundation today means reducing future risks and competing with greater confidence. We at SHM Studio we support companies on this journey, from digital strategy to web design, and responsible integration of AI tools. To delve deeper, it is possible contact us directly to explore other articles on the our blog.