The Reporting Gap: Why the Most Dangerous AI Risks Stay Quiet

Most AI governance rules assume harm announces itself. This week's posts show that assumption failing in several places at once.

Start with the EU AI Act. The serious-incident reporting duty kicks in by December 2027, but the reporting duty will fail for silent risks. Environmental harm and capability failures rarely make the news, so they rarely get reported. A duty that depends on someone noticing is a duty with holes in it.

The same deadline problem shows up in compliance work. Teams treat governance as paperwork and keep security testing in a separate room. That split will miss the 2027 deadline, because Annex III compliance needs evidence from testing, not just documents.

Enforcement has its own gap. Banned systems still appear on the market before enforcement begins, which leaves buyers holding legal and operational risk they didn't sign up for.

Outside the EU, the pattern repeats. E-23 will push Canadian banks toward fewer, bigger models as independent validation extends to AI. Consolidating many small models into a few large frozen ones may concentrate the risk the rule was written to manage.

Then there's fraud, where fake content is already doing real damage. Scammers used AI-generated images to impersonate policyholders in insurance scams, exposing weak identity checks. And a BBC report described scammers generating fake images of missing pets to extract ransom from distressed owners. Both cases turn on the same blind spot: we can't reliably verify what we're looking at.

Data privacy has its own version. A lawsuit claims Google AI in a school contributed to a teen's mental health crisis. Whatever the legal outcome, it points to concrete controls around data access and monitoring in sensitive settings.

The throughline: our governance systems are built to react to visible events. The risks that matter most are the ones no one reports, no one tests for, and no one can easily verify. Closing that gap means building detection and verification in before the deadlines arrive, not after the first quiet failure.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.