What were the actual technical vulnerabilities in the reported AI incidents?

According to the analysis, they included unverified token signatures, plaintext database connection strings and signing keys in config files, and live credentials embedded in public JavaScript—all conventional issues, not AI-specific.

How did the RIZAP leak come to light?

It was discovered through the employee's own report, not by technical detection. The number of affected people and the service name were disclosed by the client health insurance association, not by RIZAP itself.

Researchers found evidence of real-world exploitation of Microsoft 365 Copilot flaws like EchoLeak.

No real-world damage reports were confirmed within the scope of the research, and all the vulnerabilities were already fixed.

AI Safety & AlignmentZenn AI/ML3h agoOct 18, 2026, 22:00 JST

Real Invent CTO: AI-linked leaks stem from classic holes, not new AI flaws

Send on LINE

3 Key Points

What happened
The Real Invent CTO grouped reported AI-linked leaks into three types: the model itself broke in, AI became an attacker's tool, or data leaked from the user side. The technical failures—unverified tokens, plaintext DB connection strings, live credentials in public JavaScript—were conventional, not AI-specific.
Why it matters
The exposure didn't come from a new AI vulnerability class but from longstanding configuration gaps, suggesting basic hygiene and disclosure practices—not incident counts—should guide AI vendor selection, by inference.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The commentary draws on OpenAI's alignment disclosure page, which lists 19 reports and 3 notices, and on Anthropic's account of a Claude Mythos 5 incident during a CTF exercise, to show that even model developers' own testing can expose data through mundane gaps. It also cites the RIZAP case, where an employee entered data on 19,996 people into an unapproved personal AI service, to illustrate the user-side leakage category. The common thread across these incidents, in the author's reading, is not novel attack methods but the speed and breadth with which autonomous tools can probe existing weaknesses—such as the September PaperCut attack where hundreds of agents reached 395 organizations in 48 countries in under four hours. The piece closes by arguing that vendor selection should focus on disclosure practices rather than incident counts, since a vendor that discloses may simply be more transparent than one that does not.

FAQ

What were the actual technical vulnerabilities in the reported AI incidents?: According to the analysis, they included unverified token signatures, plaintext database connection strings and signing keys in config files, and live credentials embedded in public JavaScript—all conventional issues, not AI-specific.
How did the RIZAP leak come to light?: It was discovered through the employee's own report, not by technical detection. The number of affected people and the service name were disclosed by the client health insurance association, not by RIZAP itself.
Researchers found evidence of real-world exploitation of Microsoft 365 Copilot flaws like EchoLeak.: No real-world damage reports were confirmed within the scope of the research, and all the vulnerabilities were already fixed.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.