ENISA's Draft Advisory Says AI-Generated Statements Such as "All Checks Passed" Should in General Not Be Accepted Unless They Are Linked to Supporting Evidence

The European Union Agency for Cybersecurity, ENISA, has published a draft technical advisory on AI-assisted software development, dated September 2026 and marked version 0.4. It is a draft: its section headed "Note on the current version of the document" holds only a to-do line about public consultation. By its own account it "does not provide legal guidance or introduce a new compliance framework." Its purpose, it says, is "to describe the main components of AI-assisted software development, identify representative risks and threats, and provide a practical approach for integrating secure by design expectations into AI-assisted development activities." Parts of it follow, not the whole.

Risks and Failure Modes

AI-assisted practices "can improve productivity and generate functional code, but functional correctness should not be confused with security," it says. "AI-generated outputs may still contain significant vulnerabilities, especially if the main focus is on completing the requested task rather than satisfying secure by design requirements."

In its table of common risks and failure modes that arise "even where no attacker is present," one row reads: "The generated code works for the requested task but lacks relevant security controls or introduces vulnerabilities" in the system. In another, the system suggests APIs, packages, commands, configurations or design patterns that do not exist, or are outdated. In a third, developers rely on convincing code, explanations, pull request summaries or fixes without sufficient understanding or review. And under weak tests, review or evidence, "Generated tests, review notes or evidence are superficial, non-existent, incomplete, or not linked to security requirements."

Claims Without Evidence

Another table covers governance and assurance risks. One is weak traceability, where it is difficult to link AI-assisted changes to requirements, design decisions, security requirements, review outcomes, tests or evidence. Another is unsupported security claims: AI-generated statements such as "this is secure," "the issue is fixed" or "all checks passed" are "accepted without supporting evidence."

Among its good practices, organizations should determine what review or approval is needed, and what verification is required, before AI-assisted outputs are accepted, in proportion to the risk of the activity. Where appropriate, it says, evidence should be generated and retained to show what was changed and why, which checks were performed, what the results were, and who reviewed or approved the changes.

An assistant or agent can be told to help collect, summarize or format that evidence. "However, it should not replace the underlying checks or the human decision to accept the change." And, in general, ENISA says, AI-generated statements such as that "the package is safe" or that "all checks passed" "should not be accepted unless they are linked to supporting evidence."

A line that says all checks passed is output too.

When an agent reports that every check passed, where is the record of which checks ran, and what each one found?