As AI agents move from experimental sandboxes to production workflows, the consequences of a rogue action are becoming legally and operationally significant. A new runbook published on Medium by Mir Arshad Talpur provides a structured approach to handling the first 72 hours of an agent incident, leveraging ZizkaDB, an open-source causal audit database, to maintain tamper-evident records. The guide addresses a critical gap in the current infrastructure stack: the lack of standardized protocols for when an agent sends the wrong message or calls a compromised tool.

Regulatory Clocks and the EU Cyber Resilience Act

The urgency of this runbook is driven by tightening regulatory frameworks in Europe. The EU Cyber Resilience Act (CRA) reporting obligations under Article 14 began applying to manufacturers of digital products on 11 September 2026. This regime requires an early warning within 24 hours of becoming aware of a severe incident or actively exploited vulnerability, followed by a fuller notification within 72 hours. Additionally, the EU AI Act Article 73 mandates that providers of high-risk AI systems report serious incidents to market surveillance authorities immediately after establishing a causal link, with a hard deadline of 15 days from awareness. The runbook emphasizes that the clock starts at 'awareness,' not when the root cause is fully analyzed, making rapid record-keeping essential for compliance.

Phase 0: Preparing the Audit Layer

Before an incident occurs, the guide outlines a 'Phase 0' preparation strategy centered on data integrity. Teams are advised to ensure that every agent action is logged with specific metadata: the input received, context retrieved, model and prompt versions, and the specific tool called. A critical technical requirement is storing the SHA-256 checksum chain head outside the database itself. This prevents administrators from silently editing history, ensuring that the record remains tamper-evident for regulators or customers. External inputs, such as web pages or partner feeds, must be tagged with trust indicators to facilitate later tracing of poisoned data.

The First Hour: Containment and Evidence Freezing

The runbook breaks the response timeline into distinct phases. In the first hour, the priority is declaring the incident and freezing evidence. The guide provides Python code snippets for using ZizkaDB to log the incident declaration, including the 'aware_at' timestamp, which may be disputed later. Containment actions, such as disabling a tool or revoking credentials, must be logged as events linked to the incident. Crucially, the team must export the relevant events and verify the chain before any investigation begins. This step protects against the suspicion that the company altered logs to hide incompetence or liability.

Reconstructing Causality with Lineage

Between hours 1 and 8, the focus shifts to reconstruction using causal lineage. The guide suggests starting from the bad outcome and walking backwards through the event chain to identify the root cause. This process answers four key questions: what did the agent do, where did the bad input come from, who else is affected, and is the issue ongoing. By using the impact route in ZizkaDB, teams can generate a table of affected cases, which is described as the most valuable artifact of the first eight hours. This precise scoping allows for targeted customer communications rather than vague, blanket notices.

Common Pitfalls in Agent Incident Response

The author highlights several common mistakes that can turn a manageable incident into a regulatory nightmare. These include investigating before freezing evidence, failing to record the moment of awareness, and skipping containment because the cause is unclear. The guide also warns against 'overclaiming' in reports, advising teams to clearly state what is known and what remains under investigation. A practical drill is recommended for teams: pick a past case, give a colleague who wasn't involved only access to the record, and time how long it takes them to answer the four critical questions. If they can't do it in an afternoon, your instrumentation is likely insufficient.

Key Takeaways

  • Regulatory deadlines under the EU CRA and AI Act are strict and start at the moment of awareness, requiring immediate, tamper-evident logging.
  • ZizkaDB serves as a tool for maintaining causal lineage, allowing teams to trace bad outcomes back to specific inputs, tools, or versions.
  • Evidence freezingβ€”exporting the chain and storing the head externallyβ€”must happen before investigation to prevent accusations of data manipulation.
  • The runbook emphasizes that precise scoping of affected customers via impact routes is more effective and trustworthy than broad, vague notifications.

The Bottom Line

ZizkaDB is not a compliance shield, but it provides the evidentiary backbone that turns a chaotic incident response into a defensible, structured process. In an era where regulatory clocks start at awareness, the ability to prove what happened without altering the record is the single most important capability for any production AI agent.