Risk · R-173 · Governance & Accountability

Tampering with or forgery of AI records

Logs, traces or decision records of an AI system are altered, deleted or fabricated, so the record no longer shows what happened and cannot be relied on as evidence.

Severity
High
Audience
Heads-of · Practitioner
Frameworks
Principle 12 · Article 12 · Article 19 · Article 26
Reviewed
2026-11-17

What it is

Records of what an AI system did, and of what people did with it, are the evidence for investigation, explanation to affected people, regulatory reporting and proof that controls operated. They are worth only as much as their integrity. A record can be changed by an attacker covering their tracks, by an insider concealing an error, by a compromised component writing false entries, or by an agent with access to its own logs. The risk is a record that exists and is wrong. A record that does not exist is covered in record-keeping and logging failure.

How it happens

  • The system that produces records can also edit or delete them.
  • The administrators of the AI platform also administer its logs.
  • An agent can write to the store that holds its own traces.
  • Records sit in ordinary files or tables with no means of detecting change.
  • An approval or review record is created after the event to fill a gap.
  • Logs are held by a supplier whose controls the organisation cannot test.

Why it matters

An investigation can reach the wrong cause, a regulator can be given false information, and a control can appear to have operated when it did not. Once the integrity of one record is in doubt, the whole body of evidence is weakened, including the records that were never touched. The EU AI Act relies on logs of high-risk systems to make their operation traceable, which assumes the logs can be trusted.

Warning signs

  • Nobody can say who is able to change or delete AI records.
  • Records carry no protection that would reveal alteration.
  • The same team runs the system and holds its audit trail.
  • Gaps or reordering in sequences of records go unexplained.

Who owns it

The system owner is accountable for the integrity of its records. Information security designs and tests their protection. Records management and internal audit rely on that protection and test it.

In practice

An agent with write access to a customer database writes its own trace to a table in the same database. After a mistaken bulk update, an engineer under pressure deletes the trace entries before the incident review, and the review concludes that a data migration was to blame. The same pattern generalises to any record held where those it describes can change it.

Mitigating controls

Primary controls

AI record-keeping and audit trail Detective
Makes records tamper-evident in proportion to the system's tier. Inputs, outputs, model and prompt versions, decisions and human interventions are recorded automatically and retained for a defined period, so that any past event can be reconstructed.

Supporting controls

Agent action tracing Detective
Keeps agent traces out of reach of the agent and its operators. Every agent action is recorded with the instruction, plan, tool call, result and identity that produced it, in a form that allows the sequence to be reconstructed.
Independent assurance and internal audit of AI Detective
Tests that records are complete and have not been altered. Second-line review and internal audit cover AI governance and the highest-tier systems on a risk-based plan, test systems as well as documents, and track their findings to closure.

Framework and clause references

FrameworkClauseTitle
Code of Practice for the Cyber Security of AIPrinciple 12Monitor your system's behaviour Monitoring depends on logs that can be trusted.
EU AI ActArticle 12Record-keeping Logging is meant to make a high-risk system's operation traceable, which altered logs defeat.
EU AI ActArticle 19Automatically generated logs Providers must keep the logs their systems generate.
EU AI ActArticle 26Obligations of deployers of high-risk AI systems Deployers must keep the logs under their control.

Sources

EU AI Act: Regulation (EU) 2024/1689 European Union (EUR-Lex) · official text

Related terms

How to cite

Axiom Verity (2026). Tampering with or forgery of AI records. Reference Library, R-173. https://axiom-verity.com/risks/tampering-with-or-forgery-of-ai-records. Reviewed 2026-11-17.