Understanding the EU AI Act's Logging Requirements for AI Agents

The EU AI Act, spanning 144 pages, presents comprehensive logging requirements for developers of AI agents across four interrelated articles. This article examines the essential points, deadlines, and potential gaps within the regulations.

Your Agent is Likely High-Risk

Although the EU AI Act does not specifically mention 'AI agents', the classification of high-risk systems depends on their functions. Agents that engage in credit scoring, resume filtering, healthcare decision-making, insurance pricing, or emergency call triaging fall under Annex III and are categorized as high-risk.

Article 6(3) offers a potential exemption: if a system does not materially influence decision outcomes, it may escape high-risk classification. However, it is challenging to demonstrate this for agents that autonomously execute actions based on their assessments.

General-purpose AI models encounter distinct obligations under Chapter V. While the model itself may not be classified as high-risk, the system utilizing it in a high-risk context must comply with high-risk provider obligations as outlined in Article 25.

The Four Key Articles

Article 12 mandates that high-risk AI systems must enable the automatic recording of events (logs) throughout their operational lifetime. The term 'automatic' indicates the need for self-generating logs, rendering manual documentation insufficient. 'Lifetime' refers to the period from deployment to decommissioning, not limited to the current version.

According to Article 12(2), logs must encompass three categories: instances of risk or significant modifications, data for post-market monitoring, and information for operational monitoring by deployers. The regulation does not prescribe a specific format or required fields, focusing solely on these three purposes.

Article 13 stipulates that information on how deployers can collect and interpret logs must be documented. This should be viewed as a technical guide for integrating the logging layer rather than a compliance handbook.

Articles 19 and 26 establish a six-month minimum retention period for logs. Financial service providers may integrate AI logs into existing regulatory documentation, while other sectors are required to retain logs for at least six months, possibly longer, depending on specific sector regulations.

The Limitations of Standard Logs

AI agents execute tasks by calling tools, delegating responsibilities to sub-agents, obtaining responses from language models, and generating final outputs. Standard application logging can effectively capture these activities.

The challenge arises when regulators request proof that logs have not been tampered with six months later. Application logs reside on infrastructure managed by individuals who may alter or replace them without detection.

Although Article 12 does not explicitly mention 'tamper-proof' logs, if logs can be modified silently without evidence, they lose their evidentiary significance—particularly problematic for high-risk systems.

This vulnerability prompted the exploration of cryptographic signing for agent logs. This approach, which I have been developing within a project called Asqav, involves signing each agent action with a key that the agent does not possess, chaining signatures to previous actions, and storing the receipts in an untouchable location. Any alteration of an entry visibly disrupts the chain.

More importantly, the pattern of this approach is crucial. The signing key is kept outside the agent's trust boundary, every action generates a receipt, and these receipts create a verifiable chain. Whether employing NIST FIPS 204 post-quantum signatures or other methods, any implementation adhering to these principles will fulfill the intent of Article 12.

No Standard Established

Currently, there is no finalized technical standard for logging under Article 12. However, two drafts are notable: prEN 18229-1, which addresses AI logging and human oversight, and ISO/IEC DIS 24970, focused on AI system logging. Neither draft has been completed.

Organizations are preparing for regulations that specify outcomes without detailing methods. Teams that proactively establish effective logging mechanisms will be better positioned when standards are finalized, while those who delay may face challenges in adapting under pressure.

Deadlines and Penalties

Obligations outlined in Annex III take effect on August 2, 2026. Although the Commission proposed a delay through the Digital Omnibus package last November, potentially extending the deadline to December 2027, no legislation has yet been passed, making the August 2026 date enforceable.

Failure to comply may result in penalties of up to 15 million euros or 3% of global annual turnover, whichever is higher. While this formula applies to all entities, Article 99 requires penalties to be proportionate and dissuasive, instructing national authorities to consider company size and economic viability. This means startups and SMEs may face lower fines than the maximum amount, despite the formula remaining unchanged.

Key Considerations

• Can your system automatically generate logs at every decision point?
• Can those logs withstand tampering?
• Can you retain them for six months in a format that regulators can access?

If your answer is no, the August deadline may be closer than anticipated.

Source: Help Net Security News