Incident databases are lying to you about AI risk

What most people think

Most risk managers read incident databases the way they read a weather report. If a category shows zero stories, they assume the sky is clear. Environmental harm shows zero stories in the last 180 days. Lack of capability or robustness shows zero stories. Overreliance shows zero stories. So the conclusion is that these risks are negligible. That is wrong.

What the data shows

Our live incident database, which tracks reported AI failures from public news, ingested 855 stories in the last 180 days. The categories with the most attention are privacy leaks, fraud, security vulnerabilities, and multi-agent risks. Each has dozens of stories. The categories with zero stories are environmental harm, robustness, and overreliance.

The top five reporting outlets carry only 9% of recent stories. That means coverage is fragmented and driven by what sells: dramatic breaches and compliance failures. Nobody clicks on a headline about a model slowly getting worse at its job or a data center quietly burning more power.

Meanwhile, the security side of the industry is writing about things that never reach the governance feed. Our sister platform ThreatClaw published 40 articles in 60 days on AI security topics: critical vulnerabilities, actively exploited flaws, supply chain attacks. The titles are concrete: "Attackers Can Quietly Rewrite an AI's Memory," "Your 'Deleted' Data Still Lives Inside Your AI." None of those would make it into a mainstream incident database as a robustness or environmental story.

Why this happens

Incident databases are built from news reports. News reports are built from what editors think readers will click. A story about a privacy leak is easy to understand and alarming. A story about a model that drifts off-spec over eighteen months is neither. So it never gets written, and it never gets counted.

This is a structural bias, not a measurement of reality. The absence of stories in a category tells you that the category is not newsworthy, not that it is safe.

There is also a timing problem. Robustness failures are slow. They compound through model drift, hardware degradation, and changing data distributions. By the time they cause a visible incident, the root cause is months old and hard to trace to a single moment. News cycles do not work that way.

Environmental harm is even more invisible. It is a slow accumulation of compute waste and carbon output. No single event triggers a headline. It only becomes a problem when a regulator asks for the numbers.

The best argument against this

You could argue that zero stories means the risk is genuinely low. If no one has been harmed by a robustness failure in six months, maybe the risk is not real. And environmental harm is a cost, not an incident. It does not cause immediate damage to users.

That argument fails because absence of evidence is not evidence of absence. The MIT AI Risk Repository, a catalogued taxonomy of AI risks, lists robustness and environmental harm as real subdomains. The MITRE ATLAS framework documents 22 real-world cases of prompt crafting attacks and 18 cases of model evasion. Those are documented, not hypothetical. The news just does not report them.

Also, the legal timeline is moving. The EU AI Act's high-risk rules apply from December 2027, and its rules on embedded high-risk systems apply from August 2028. Those rules require you to demonstrate robustness and to report serious incidents. If you have no data on robustness because your monitoring tools only track news, you will fail the assessment.

What I think happens next

By August 2028, when the EU AI Act's Annex I obligations on embedded high-risk systems apply, at least two major industrial AI deployments will face sudden regulatory halts due to unmeasured robustness failures. These failures will not have appeared in any prior incident feed because no one was measuring them.

What would prove me wrong: if all regulatory halts or enforcement actions under the EU AI Act before August 2028 come exclusively from risks that were already tracked in the top-five incident categories. If that happens, the news-driven taxonomy was sufficient. I do not think it will.

What to do about it

Start measuring what the news does not cover. You can begin this week.

First, implement continuous automated stress-testing for model robustness. Run your models against edge cases and distribution shifts on a schedule. Log the results. Do not wait for a news story to tell you something is wrong.

Second, establish internal telemetry for compute resource consumption and carbon output. Track energy use per inference and per training run. This gives you the data you need for environmental compliance obligations under the EU AI Act and other frameworks.

Third, stop relying solely on external incident databases for your risk register. Use them as one input, not the whole picture. Build your own operational metrics for the categories that matter to your deployment.

Fourth, review the MITRE ATLAS attack techniques that have documented real-world cases, such as LLM prompt crafting and model evasion. Map them to your systems and check whether you have monitoring in place.

Fifth, if you want to see what an AI actually did versus what it was allowed to do, tools like Argus (argus.threatclaw.ai) can record every trace an AI application produces and scan for hidden attacks. That is not a compliance solution, but it gives you the visibility you need to catch the failures that never make the news.

A governance programme that only watches the news is not a governance programme. It is a clipping service. The risks that will actually hurt you are the ones that never get a headline.

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.

Related reading:

• How Whistleblower Reports on AI Failures Get Hijacked Before Anyone Reads Them on ThreatClaw
• When Your AI Fails, Who Do You Tell? The New Rules Are Here on ThreatClaw