GPAI Models With Systemic Risk: The Safety and Security Chapter of the Code of Practice, Explained

General-purpose AI (GPAI) providers all owe transparency and copyright duties, but the most capable models carry a heavier layer: managing systemic risk. The Safety and Security chapter of the GPAI Code of Practice is the Commission-endorsed route for showing you do that. With the AI Office able to enforce since 2 August 2026, this chapter is no longer theory.

Two tiers of GPAI obligations

All GPAI providers must keep technical documentation, share information with downstream providers, publish a training data summary and maintain a copyright policy. Our post on the Copyright chapter covers that layer. Providers of models with systemic risk add risk assessment, evaluation, incident reporting and cybersecurity duties. The Act presumes systemic risk above a training-compute threshold of 10^25 floating-point operations; confirm the current threshold and any Commission guidance, because the BCLP summary I used does not state a figure.

What the Safety and Security chapter requires

According to BCLP's key takeaways, signatories should:

Example mitigations include filtering and cleaning training data, changing model behaviour, staging access to the model and providing tools to deployers. Governance should allocate responsibility for systemic risk clearly, with oversight at senior management level where appropriate.

Serious incident notification

Serious incidents must be reported to the AI Office on tight deadlines. The chapter summary gives two days for serious and irreversible disruption of critical infrastructure and five days for a serious cybersecurity breach, with providers of systemic-risk models expected to notify the AI Office without delay of relevant developments per BCLP. Verify the exact timelines against the final Code text and Article 55 before building your playbook. If you also ship software into the EU, align this with the Cyber Resilience Act reporting duties.

Key dates

The AI Office will review the Code at least every two years. The fine ceiling is covered in our fines explainer.

Signing the Code versus complying without it

Adhering to the Code gives a clearer route to demonstrating compliance and, in practice, a lighter information burden. Providers that do not sign must show equivalent compliance by other adequate means. The first 60 days of enforcement show the AI Office already sending information requests to GPAI providers, so unsigned providers should expect more questions.

Open-source models

Open-source GPAI models get partial relief, but not if they present systemic risk. See our open-source exemption guide for the boundaries.

What downstream users should do

Companies building on a systemic-risk model should ask for the provider's documentation, confirm whether it signed the Code, and map the model's limits into their own risk file. Our agentic AI guide shows how model-level and system-level duties combine.

Bottom line

For systemic-risk providers, the chapter turns into an annual cycle: assess, mitigate, document, report. Build the framework and the incident path now; the 2027 legacy deadline arrives sooner than a new governance process takes to mature.