⭐ Star on GitHub

Summary: The most prescriptive security control set in this corpus. Five control families protecting unreleased model weights: general hygiene, parameter protection, interface hardening, insider threat, and security assurance.

Sources: raw/Laws/Overview of the Code of Practice _ EU Artificial Intelligence Act.md (Commitment 6, Appendix 4)

Last updated: 2026-07-28


Scope and the Security Goal

Commitment 6 — signatories “maintain robust cybersecurity measures throughout the model lifecycle to prevent risks arising from unauthorized release, access, or theft.”

Two scoping rules (source: Overview of the Code of Practice…md):

  • Excludes “models with capabilities lower than at least one model with parameters available for public download.” If an open-weight model already matches your capabilities, protecting yours protects nothing.
  • Measures “remain in effect until model parameters are made publicly available or securely deleted.”

Measure 6.1 — Security Goal: signatories “shall establish a comprehensive Security Goal that identifies threat actors their mitigations are intended to protect against, informed by current and projected model capabilities.”

This is a threat-model-first requirement, and the “projected model capabilities” clause means the goal has to be set against where the model is going, not where it is.

Measure 6.2 — deviations from Appendix 4.1–4.5(a) “must demonstrate equivalent protective outcomes.” Implementation “may be phased to correspond with model capability advancement.”

Appendix 4.1 — General security mitigations

Baseline hygiene, unremarkable but stated:

  • Network access control — “Identity and access management (e.g. MFA, strong passwords, zero trust architecture, wireless security equal to wired, guest network isolation).”
  • Social engineering protection — “Email filtering to detect phishing and suspicious attachments.”
  • Malware and removable media — “Policies restricting the use of USBs and similar devices.”
  • Software security — “Regular software updates and patch management.”

Appendix 4.2 — Protection of unreleased model parameters

This is the distinctive part — weights treated as crown-jewel material:

  • Track all stored copies — “Maintain a secure registry of devices/locations holding model parameters.”
  • Restrict copying to unmanaged devices — access controls plus monitoring for unauthorised transfer.
  • Encrypt in transit and at rest — “Use 256-bit encryption and secure key storage (e.g. TPM).”
  • Secure temporary storage — “Decrypt parameters only in non-persistent memory for legitimate use.”
  • Secure parameters in use — “Deploy confidential computing techniques such as attested trusted execution environments.”
  • Control physical access — limit access to data centres, “perform inspections for unauthorised presence or devices.”

The “secure parameters in use” requirement is the demanding one. Encrypting at rest and in transit is standard; requiring attested TEEs for weights in use closes the memory-dump path and is a materially higher bar than most production ML infrastructure meets.

Appendix 4.3 — Hardening interface access

  • Limit interface access — authorised users/software only, MFA, “review permissions at least every 6 months.”
  • Secure interface code — “in-depth manual or automated security reviews of code linked to model parameter access.”
  • Prevent exfiltration — “Apply methods such as output rate limiting on interfaces.”
  • Minimise insider access — limit who can reach non-hardened interfaces.

Output rate limiting appears here as an anti-exfiltration control. It is the same control that limits T09 gradient-based attacks and best-of-N search, which “require… the ability to send thousands of queries.”

Appendix 4.4 — Insider threats

Explicitly scoped to sabotage or theft “including by or through models”:

  • Personnel vetting — background checks for those with access to sensitive model data/systems.
  • Insider threat awareness — train staff to recognise and report.
  • Prevent self-exfiltration by models — “Use sandboxing and code execution isolation.”
  • Safeguard model training — “Inspect training data for tampering or sabotage.”

Two of these four are unusual in a security baseline. “Prevent self-exfiltration by models” treats the model itself as a potential insider threat — the concrete control against the “autonomy, self-replication” and “self-reasoning and evasion of oversight” capabilities listed as systemic risk sources. “Inspect training data for tampering” is the operational answer to Art 15(5) data poisoning.

Appendix 4.5 — Security assurance

  • Independent external reviews “if internal capacity is insufficient.”
  • Red-teaming “to identify security gaps in networks and facilities.”
  • Bug bounty programs for public-facing endpoints “where appropriate.”
  • Test insider mitigation protocols, “including personnel integrity assessments.”
  • Secure third-party channels for issue reporting.
  • Active monitoring with “Endpoint Detection and Response (EDR) or Intrusion Detection Systems (IDS).”

Note the red teaming here is infrastructure red teaming — networks and facilities — distinct from the model adversarial testing required under Commitment 3 and Art 55(1)(a). Both are required; they are different exercises with different teams.

What this control set is actually defending

Read as a whole, Appendix 4 defends against weight theft, not against misuse of the deployed model. Registry, encryption, TEEs, interface hardening, insider vetting, physical inspection — the asset is the parameter file.

That is coherent with the threat model: once weights leak, abliteration removes safety behaviour entirely and no model-level control survives. Weight confidentiality is the precondition for every other safety measure holding.

The corresponding model-behaviour controls live in Commitment 5 (safety mitigations): “filtering, continuous monitoring, refusal training, phased access controls, downstream tool safeguards, and secure deployment environments” (source: Overview of the Code of Practice…md).

Gap worth noting

Appendix 4 has nothing on prompt injection, tool poisoning or agentic containment beyond “sandboxing and code execution isolation” for model self-exfiltration. The agentic attack surface is acknowledged in the systemic-risk taxonomy (“tool use and control of physical systems”) but not in the control set. For an organisation deploying agents, Appendix 4 is necessary and nowhere near sufficient.