AI Incident Reports Are Written by Lawyers and Security Vendors, Not by Reality

What most people think

Most people in AI governance assume that incident reporting is a neutral, comprehensive record of what goes wrong with AI systems. They think the media coverage reflects the true frequency and severity of incidents. If a risk category gets little coverage, they assume it is rare. If it gets a lot, they assume it is common. That assumption is wrong.

What the data shows

Our live incident database, which tracks reported AI failures from public news, shows a clear pattern. In the last 180 days, we logged 762 stories. The top five outlets carried 9% of the last 45 days' stories. Those outlets are JD Supra, Help Net Security, Law360, Tech Policy Press, and Biometric Update. Three of the five are legal or security trade publications.

The skew is not just in the outlets. It is in the categories. Using the MIT AI Risk Repository's subdomains, we matched stories to risk classes. Overreliance and unsafe use (5.1) had zero stories in 180 days. Environmental harm (6.6) had zero. Lack of capability or robustness (7.3) had zero. Competitive dynamics (6.4) had one. Meanwhile, privacy (2.1) had 76 stories, fraud (4.3) had 58, multi-agent risks (7.6) had 55, and security vulnerabilities (2.2) had 53.

The same pattern appears in attack techniques. MITRE ATLAS documents real-world cases for techniques like LLM prompt crafting (22 cases) and AI agent tool invocation (15 cases). The news window never mentions them.

Why this happens

Legal outlets cover incidents with regulatory or litigation implications. Privacy breaches, fraud, and security vulnerabilities have clear legal hooks. They trigger lawsuits, fines, and regulatory actions. So they get covered. Overreliance and environmental harm do not have such hooks. No one sues over an AI that makes users overly dependent. No one files a class action over a model's carbon footprint. So they get zero coverage.

Security outlets cover technical vulnerabilities. They are in the business of selling threat intelligence and detection tools. Their incentive is to highlight risks that fit their product categories. That is why security stories dominate the incident database. The result is a feedback loop. Risk teams read the trade press, so they prioritize the risks that get headlines. They under-invest in risks that do not generate legal or security stories, even when those risks are real and consequential.

The best argument against this

One could argue that the trade press is simply reflecting what is actually happening. Maybe overreliance and environmental harm are genuinely rare, and the lack of coverage is accurate. The legal and security categories get attention because they are the ones that actually occur.

That argument fails on two counts. First, the MITRE ATLAS data shows that documented cases exist for under-reported techniques. They are not rare; they are just not newsworthy by the trade press's standards. Second, the pattern is not about frequency. It is about what has a legal or security angle. A serious incident that harms users through overreliance may not be a privacy breach or a security vulnerability, so it never makes the news. That does not mean it is less severe. It means it is less visible.

What I think happens next

By June 2026, at least one major AI incident in an under-reported category, such as environmental harm or overreliance, will be revealed through a non-traditional source. It will come from an academic paper, a whistleblower, or a citizen report. The trade press will have missed it entirely. That will be the moment the bias becomes undeniable.

What would prove me wrong: no such incident emerges, and the trade press continues to be the primary source for all consequential AI incidents. If that happens, I will revise my view.

What to do about it

Start this week. Here are five concrete steps.

1. Diversify your incident intelligence sources. Do not rely only on trade press. Add academic research, whistleblower platforms, and regulatory filings. Our own database pulls from public news, but you should go beyond that.

2. Run a horizon scan for the zero-coverage risk categories. Specifically look for early warning signs of overreliance and environmental harm. These are the risks that will blindside you.

3. Build a monitoring list for academic papers and conference proceedings. Many AI failures are documented in research before they become public incidents.

4. Set up alerts for regulatory filings and court dockets. These often reveal incidents that never make the news.

5. Use the MIT AI Risk Repository and MITRE ATLAS as structured taxonomies. Map your own risk register against them to see what you are missing.

You can also read a related piece on ThreatClaw about who to tell when your AI fails: https://www.threatclaw.ai/blog/when-your-ai-fails-who-do-you-tell-the-new-rules-are-here. It covers the new reporting rules under the EU AI Act, which start in December 2027. But do not wait for the rules. Start fixing your incident intelligence now.

The EU AI Act's high-risk rules apply from December 2027, and its serious-incident reporting requirement starts the same day. Colorado's AI law, which was repealed and replaced, takes effect January 2027. Canada has no comprehensive federal AI law. The legal landscape is shifting, but the bias in incident reporting is not going to fix itself. You have to fix it.

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:

• When Your AI Fails, Who Do You Tell? The New Rules Are Here on ThreatClaw
• Hackers Read the Government's 'Must Patch' List Too on ThreatClaw